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_locked(struct fsnotify_mark *mark, 125 struct fsnotify_group *group) 126 { 127 struct inode *inode = NULL; 128 129 BUG_ON(!mutex_is_locked(&group->mark_mutex)); 130 131 spin_lock(&mark->lock); 132 133 /* something else already called this function on this mark */ 134 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) { 135 spin_unlock(&mark->lock); 136 return; 137 } 138 139 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE; 140 141 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) { 142 inode = mark->i.inode; 143 fsnotify_destroy_inode_mark(mark); 144 } else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT) 145 fsnotify_destroy_vfsmount_mark(mark); 146 else 147 BUG(); 148 149 list_del_init(&mark->g_list); 150 151 spin_unlock(&mark->lock); 152 153 if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED)) 154 iput(inode); 155 /* release lock temporarily */ 156 mutex_unlock(&group->mark_mutex); 157 158 spin_lock(&destroy_lock); 159 list_add(&mark->destroy_list, &destroy_list); 160 spin_unlock(&destroy_lock); 161 wake_up(&destroy_waitq); 162 /* 163 * We don't necessarily have a ref on mark from caller so the above destroy 164 * may have actually freed it, unless this group provides a 'freeing_mark' 165 * function which must be holding a reference. 166 */ 167 168 /* 169 * Some groups like to know that marks are being freed. This is a 170 * callback to the group function to let it know that this mark 171 * is being freed. 172 */ 173 if (group->ops->freeing_mark) 174 group->ops->freeing_mark(mark, group); 175 176 /* 177 * __fsnotify_update_child_dentry_flags(inode); 178 * 179 * I really want to call that, but we can't, we have no idea if the inode 180 * still exists the second we drop the mark->lock. 181 * 182 * The next time an event arrive to this inode from one of it's children 183 * __fsnotify_parent will see that the inode doesn't care about it's 184 * children and will update all of these flags then. So really this 185 * is just a lazy update (and could be a perf win...) 186 */ 187 188 atomic_dec(&group->num_marks); 189 190 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING); 191 } 192 193 void fsnotify_destroy_mark(struct fsnotify_mark *mark, 194 struct fsnotify_group *group) 195 { 196 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING); 197 fsnotify_destroy_mark_locked(mark, group); 198 mutex_unlock(&group->mark_mutex); 199 } 200 201 void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask) 202 { 203 assert_spin_locked(&mark->lock); 204 205 mark->mask = mask; 206 207 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) 208 fsnotify_set_inode_mark_mask_locked(mark, mask); 209 } 210 211 void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask) 212 { 213 assert_spin_locked(&mark->lock); 214 215 mark->ignored_mask = mask; 216 } 217 218 /* 219 * Attach an initialized mark to a given group and fs object. 220 * These marks may be used for the fsnotify backend to determine which 221 * event types should be delivered to which group. 222 */ 223 int fsnotify_add_mark_locked(struct fsnotify_mark *mark, 224 struct fsnotify_group *group, struct inode *inode, 225 struct vfsmount *mnt, int allow_dups) 226 { 227 int ret = 0; 228 229 BUG_ON(inode && mnt); 230 BUG_ON(!inode && !mnt); 231 BUG_ON(!mutex_is_locked(&group->mark_mutex)); 232 233 /* 234 * LOCKING ORDER!!!! 235 * group->mark_mutex 236 * mark->lock 237 * inode->i_lock 238 */ 239 spin_lock(&mark->lock); 240 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE; 241 242 fsnotify_get_group(group); 243 mark->group = group; 244 list_add(&mark->g_list, &group->marks_list); 245 atomic_inc(&group->num_marks); 246 fsnotify_get_mark(mark); /* for i_list and g_list */ 247 248 if (inode) { 249 ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups); 250 if (ret) 251 goto err; 252 } else if (mnt) { 253 ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups); 254 if (ret) 255 goto err; 256 } else { 257 BUG(); 258 } 259 260 /* this will pin the object if appropriate */ 261 fsnotify_set_mark_mask_locked(mark, mark->mask); 262 spin_unlock(&mark->lock); 263 264 if (inode) 265 __fsnotify_update_child_dentry_flags(inode); 266 267 return ret; 268 err: 269 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE; 270 list_del_init(&mark->g_list); 271 fsnotify_put_group(group); 272 mark->group = NULL; 273 atomic_dec(&group->num_marks); 274 275 spin_unlock(&mark->lock); 276 277 spin_lock(&destroy_lock); 278 list_add(&mark->destroy_list, &destroy_list); 279 spin_unlock(&destroy_lock); 280 wake_up(&destroy_waitq); 281 282 return ret; 283 } 284 285 int fsnotify_add_mark(struct fsnotify_mark *mark, struct fsnotify_group *group, 286 struct inode *inode, struct vfsmount *mnt, int allow_dups) 287 { 288 int ret; 289 mutex_lock(&group->mark_mutex); 290 ret = fsnotify_add_mark_locked(mark, group, inode, mnt, allow_dups); 291 mutex_unlock(&group->mark_mutex); 292 return ret; 293 } 294 295 /* 296 * clear any marks in a group in which mark->flags & flags is true 297 */ 298 void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group, 299 unsigned int flags) 300 { 301 struct fsnotify_mark *lmark, *mark; 302 303 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING); 304 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) { 305 if (mark->flags & flags) { 306 fsnotify_get_mark(mark); 307 fsnotify_destroy_mark_locked(mark, group); 308 fsnotify_put_mark(mark); 309 } 310 } 311 mutex_unlock(&group->mark_mutex); 312 } 313 314 /* 315 * Given a group, destroy all of the marks associated with that group. 316 */ 317 void fsnotify_clear_marks_by_group(struct fsnotify_group *group) 318 { 319 fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1); 320 } 321 322 void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old) 323 { 324 assert_spin_locked(&old->lock); 325 new->i.inode = old->i.inode; 326 new->m.mnt = old->m.mnt; 327 if (old->group) 328 fsnotify_get_group(old->group); 329 new->group = old->group; 330 new->mask = old->mask; 331 new->free_mark = old->free_mark; 332 } 333 334 /* 335 * Nothing fancy, just initialize lists and locks and counters. 336 */ 337 void fsnotify_init_mark(struct fsnotify_mark *mark, 338 void (*free_mark)(struct fsnotify_mark *mark)) 339 { 340 memset(mark, 0, sizeof(*mark)); 341 spin_lock_init(&mark->lock); 342 atomic_set(&mark->refcnt, 1); 343 mark->free_mark = free_mark; 344 } 345 346 static int fsnotify_mark_destroy(void *ignored) 347 { 348 struct fsnotify_mark *mark, *next; 349 LIST_HEAD(private_destroy_list); 350 351 for (;;) { 352 spin_lock(&destroy_lock); 353 /* exchange the list head */ 354 list_replace_init(&destroy_list, &private_destroy_list); 355 spin_unlock(&destroy_lock); 356 357 synchronize_srcu(&fsnotify_mark_srcu); 358 359 list_for_each_entry_safe(mark, next, &private_destroy_list, destroy_list) { 360 list_del_init(&mark->destroy_list); 361 fsnotify_put_mark(mark); 362 } 363 364 wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list)); 365 } 366 367 return 0; 368 } 369 370 static int __init fsnotify_mark_init(void) 371 { 372 struct task_struct *thread; 373 374 thread = kthread_run(fsnotify_mark_destroy, NULL, 375 "fsnotify_mark"); 376 if (IS_ERR(thread)) 377 panic("unable to start fsnotify mark destruction thread."); 378 379 return 0; 380 } 381 device_initcall(fsnotify_mark_init); 382