Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 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 guarenteed 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> |
Eric Paris | 75c1be4 | 2010-07-28 10:18:38 -0400 | [diff] [blame^] | 88 | #include <linux/kthread.h> |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 89 | #include <linux/module.h> |
| 90 | #include <linux/mutex.h> |
| 91 | #include <linux/slab.h> |
| 92 | #include <linux/spinlock.h> |
Eric Paris | 75c1be4 | 2010-07-28 10:18:38 -0400 | [diff] [blame^] | 93 | #include <linux/srcu.h> |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 94 | #include <linux/writeback.h> /* for inode_lock */ |
| 95 | |
| 96 | #include <asm/atomic.h> |
| 97 | |
| 98 | #include <linux/fsnotify_backend.h> |
| 99 | #include "fsnotify.h" |
| 100 | |
Eric Paris | 75c1be4 | 2010-07-28 10:18:38 -0400 | [diff] [blame^] | 101 | struct srcu_struct fsnotify_mark_srcu; |
| 102 | static DEFINE_SPINLOCK(destroy_lock); |
| 103 | static LIST_HEAD(destroy_list); |
| 104 | static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq); |
| 105 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 106 | void fsnotify_get_mark(struct fsnotify_mark *mark) |
| 107 | { |
| 108 | atomic_inc(&mark->refcnt); |
| 109 | } |
| 110 | |
| 111 | void fsnotify_put_mark(struct fsnotify_mark *mark) |
| 112 | { |
| 113 | if (atomic_dec_and_test(&mark->refcnt)) |
| 114 | mark->free_mark(mark); |
| 115 | } |
| 116 | |
| 117 | /* |
| 118 | * Any time a mark is getting freed we end up here. |
| 119 | * The caller had better be holding a reference to this mark so we don't actually |
| 120 | * do the final put under the mark->lock |
| 121 | */ |
| 122 | void fsnotify_destroy_mark(struct fsnotify_mark *mark) |
| 123 | { |
| 124 | struct fsnotify_group *group; |
Eric Paris | 0d48b7f | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 125 | struct inode *inode = NULL; |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 126 | |
| 127 | spin_lock(&mark->lock); |
| 128 | |
| 129 | group = mark->group; |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 130 | |
Eric Paris | 700307a | 2010-07-28 10:18:38 -0400 | [diff] [blame] | 131 | /* something else already called this function on this mark */ |
| 132 | if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) { |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 133 | spin_unlock(&mark->lock); |
| 134 | return; |
| 135 | } |
| 136 | |
Eric Paris | 700307a | 2010-07-28 10:18:38 -0400 | [diff] [blame] | 137 | mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE; |
| 138 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 139 | /* 1 from caller and 1 for being on i_list/g_list */ |
| 140 | BUG_ON(atomic_read(&mark->refcnt) < 2); |
| 141 | |
| 142 | spin_lock(&group->mark_lock); |
| 143 | |
Eric Paris | 0d48b7f | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 144 | if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) { |
Eric Paris | 0d48b7f | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 145 | inode = mark->i.inode; |
Eric Paris | b31d397 | 2010-04-21 16:49:38 -0400 | [diff] [blame] | 146 | fsnotify_destroy_inode_mark(mark); |
Eric Paris | 0d48b7f | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 147 | } else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT) |
| 148 | fsnotify_destroy_vfsmount_mark(mark); |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 149 | else |
| 150 | BUG(); |
| 151 | |
| 152 | list_del_init(&mark->g_list); |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 153 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 154 | spin_unlock(&group->mark_lock); |
| 155 | spin_unlock(&mark->lock); |
| 156 | |
Eric Paris | 75c1be4 | 2010-07-28 10:18:38 -0400 | [diff] [blame^] | 157 | spin_lock(&destroy_lock); |
| 158 | list_add(&mark->destroy_list, &destroy_list); |
| 159 | spin_unlock(&destroy_lock); |
| 160 | wake_up(&destroy_waitq); |
| 161 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 162 | /* |
| 163 | * Some groups like to know that marks are being freed. This is a |
| 164 | * callback to the group function to let it know that this mark |
| 165 | * is being freed. |
| 166 | */ |
| 167 | if (group->ops->freeing_mark) |
| 168 | group->ops->freeing_mark(mark, group); |
| 169 | |
| 170 | /* |
| 171 | * __fsnotify_update_child_dentry_flags(inode); |
| 172 | * |
| 173 | * I really want to call that, but we can't, we have no idea if the inode |
| 174 | * still exists the second we drop the mark->lock. |
| 175 | * |
| 176 | * The next time an event arrive to this inode from one of it's children |
| 177 | * __fsnotify_parent will see that the inode doesn't care about it's |
| 178 | * children and will update all of these flags then. So really this |
| 179 | * is just a lazy update (and could be a perf win...) |
| 180 | */ |
| 181 | |
Eric Paris | 90b1e7a | 2009-12-17 21:24:33 -0500 | [diff] [blame] | 182 | if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED)) |
Eric Paris | 0d48b7f | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 183 | iput(inode); |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 184 | |
| 185 | /* |
| 186 | * it's possible that this group tried to destroy itself, but this |
| 187 | * this mark was simultaneously being freed by inode. If that's the |
| 188 | * case, we finish freeing the group here. |
| 189 | */ |
| 190 | if (unlikely(atomic_dec_and_test(&group->num_marks))) |
| 191 | fsnotify_final_destroy_group(group); |
| 192 | } |
| 193 | |
Eric Paris | 90b1e7a | 2009-12-17 21:24:33 -0500 | [diff] [blame] | 194 | void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask) |
| 195 | { |
| 196 | assert_spin_locked(&mark->lock); |
| 197 | |
| 198 | mark->mask = mask; |
| 199 | |
| 200 | if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) |
| 201 | fsnotify_set_inode_mark_mask_locked(mark, mask); |
| 202 | } |
| 203 | |
Eric Paris | 33af5e3 | 2009-12-17 21:24:33 -0500 | [diff] [blame] | 204 | void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask) |
| 205 | { |
| 206 | assert_spin_locked(&mark->lock); |
| 207 | |
| 208 | mark->ignored_mask = mask; |
| 209 | } |
Eric Paris | 90b1e7a | 2009-12-17 21:24:33 -0500 | [diff] [blame] | 210 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 211 | /* |
| 212 | * Attach an initialized mark to a given group and fs object. |
| 213 | * These marks may be used for the fsnotify backend to determine which |
| 214 | * event types should be delivered to which group. |
| 215 | */ |
| 216 | int fsnotify_add_mark(struct fsnotify_mark *mark, |
| 217 | struct fsnotify_group *group, struct inode *inode, |
| 218 | struct vfsmount *mnt, int allow_dups) |
| 219 | { |
| 220 | int ret = 0; |
| 221 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 222 | BUG_ON(inode && mnt); |
| 223 | BUG_ON(!inode && !mnt); |
| 224 | |
| 225 | /* |
| 226 | * if this group isn't being testing for inode type events we need |
| 227 | * to start testing |
| 228 | */ |
| 229 | if (inode && unlikely(list_empty(&group->inode_group_list))) |
| 230 | fsnotify_add_inode_group(group); |
| 231 | else if (mnt && unlikely(list_empty(&group->vfsmount_group_list))) |
| 232 | fsnotify_add_vfsmount_group(group); |
| 233 | |
| 234 | /* |
| 235 | * LOCKING ORDER!!!! |
| 236 | * mark->lock |
| 237 | * group->mark_lock |
| 238 | * inode->i_lock |
| 239 | */ |
| 240 | spin_lock(&mark->lock); |
| 241 | spin_lock(&group->mark_lock); |
| 242 | |
Eric Paris | 700307a | 2010-07-28 10:18:38 -0400 | [diff] [blame] | 243 | mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE; |
| 244 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 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; |
Eric Paris | 0d48b7f | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 254 | } else if (mnt) { |
| 255 | ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups); |
| 256 | if (ret) |
| 257 | goto err; |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 258 | } else { |
| 259 | BUG(); |
| 260 | } |
| 261 | |
| 262 | spin_unlock(&group->mark_lock); |
Eric Paris | 90b1e7a | 2009-12-17 21:24:33 -0500 | [diff] [blame] | 263 | |
| 264 | /* this will pin the object if appropriate */ |
| 265 | fsnotify_set_mark_mask_locked(mark, mark->mask); |
| 266 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 267 | spin_unlock(&mark->lock); |
| 268 | |
| 269 | if (inode) |
| 270 | __fsnotify_update_child_dentry_flags(inode); |
| 271 | |
| 272 | return ret; |
| 273 | err: |
Eric Paris | 700307a | 2010-07-28 10:18:38 -0400 | [diff] [blame] | 274 | mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE; |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 275 | list_del_init(&mark->g_list); |
Eric Paris | 75c1be4 | 2010-07-28 10:18:38 -0400 | [diff] [blame^] | 276 | mark->group = NULL; |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 277 | atomic_dec(&group->num_marks); |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 278 | |
| 279 | spin_unlock(&group->mark_lock); |
| 280 | spin_unlock(&mark->lock); |
| 281 | |
Eric Paris | 75c1be4 | 2010-07-28 10:18:38 -0400 | [diff] [blame^] | 282 | spin_lock(&destroy_lock); |
| 283 | list_add(&mark->destroy_list, &destroy_list); |
| 284 | spin_unlock(&destroy_lock); |
| 285 | wake_up(&destroy_waitq); |
| 286 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 287 | return ret; |
| 288 | } |
| 289 | |
| 290 | /* |
Eric Paris | 4d92604 | 2009-12-17 21:24:34 -0500 | [diff] [blame] | 291 | * clear any marks in a group in which mark->flags & flags is true |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 292 | */ |
Eric Paris | 4d92604 | 2009-12-17 21:24:34 -0500 | [diff] [blame] | 293 | void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group, |
| 294 | unsigned int flags) |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 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) { |
Eric Paris | 4d92604 | 2009-12-17 21:24:34 -0500 | [diff] [blame] | 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 | } |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 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 | |
Eric Paris | 4d92604 | 2009-12-17 21:24:34 -0500 | [diff] [blame] | 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 | |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 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 | new->group = old->group; |
| 329 | new->mask = old->mask; |
| 330 | new->free_mark = old->free_mark; |
| 331 | } |
| 332 | |
| 333 | /* |
| 334 | * Nothing fancy, just initialize lists and locks and counters. |
| 335 | */ |
| 336 | void fsnotify_init_mark(struct fsnotify_mark *mark, |
| 337 | void (*free_mark)(struct fsnotify_mark *mark)) |
| 338 | { |
Eric Paris | ba643f0 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 339 | memset(mark, 0, sizeof(*mark)); |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 340 | spin_lock_init(&mark->lock); |
| 341 | atomic_set(&mark->refcnt, 1); |
Eric Paris | 5444e29 | 2009-12-17 21:24:27 -0500 | [diff] [blame] | 342 | mark->free_mark = free_mark; |
| 343 | } |
Eric Paris | 75c1be4 | 2010-07-28 10:18:38 -0400 | [diff] [blame^] | 344 | |
| 345 | static int fsnotify_mark_destroy(void *ignored) |
| 346 | { |
| 347 | struct fsnotify_mark *mark, *next; |
| 348 | LIST_HEAD(private_destroy_list); |
| 349 | |
| 350 | for (;;) { |
| 351 | spin_lock(&destroy_lock); |
| 352 | list_for_each_entry_safe(mark, next, &destroy_list, destroy_list) { |
| 353 | list_del(&mark->destroy_list); |
| 354 | list_add(&mark->destroy_list, &private_destroy_list); |
| 355 | } |
| 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); |