mark.c

   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 *  Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
   4 */
   5
   6/*
   7 * fsnotify inode mark locking/lifetime/and refcnting
   8 *
   9 * REFCNT:
  10 * The group->recnt and mark->refcnt tell how many "things" in the kernel
  11 * currently are referencing the objects. Both kind of objects typically will
  12 * live inside the kernel with a refcnt of 2, one for its creation and one for
  13 * the reference a group and a mark hold to each other.
  14 * If you are holding the appropriate locks, you can take a reference and the
  15 * object itself is guaranteed to survive until the reference is dropped.
  16 *
  17 * LOCKING:
  18 * There are 3 locks involved with fsnotify inode marks and they MUST be taken
  19 * in order as follows:
  20 *
  21 * group->mark_mutex
  22 * mark->lock
  23 * mark->connector->lock
  24 *
  25 * group->mark_mutex protects the marks_list anchored inside a given group and
  26 * each mark is hooked via the g_list.  It also protects the groups private
  27 * data (i.e group limits).
  28
  29 * mark->lock protects the marks attributes like its masks and flags.
  30 * Furthermore it protects the access to a reference of the group that the mark
  31 * is assigned to as well as the access to a reference of the inode/vfsmount
  32 * that is being watched by the mark.
  33 *
  34 * mark->connector->lock protects the list of marks anchored inside an
  35 * inode / vfsmount and each mark is hooked via the i_list.
  36 *
  37 * A list of notification marks relating to inode / mnt is contained in
  38 * fsnotify_mark_connector. That structure is alive as long as there are any
  39 * marks in the list and is also protected by fsnotify_mark_srcu. A mark gets
  40 * detached from fsnotify_mark_connector when last reference to the mark is
  41 * dropped.  Thus having mark reference is enough to protect mark->connector
  42 * pointer and to make sure fsnotify_mark_connector cannot disappear. Also
  43 * because we remove mark from g_list before dropping mark reference associated
  44 * with that, any mark found through g_list is guaranteed to have
  45 * mark->connector set until we drop group->mark_mutex.
  46 *
  47 * LIFETIME:
  48 * Inode marks survive between when they are added to an inode and when their
  49 * refcnt==0. Marks are also protected by fsnotify_mark_srcu.
  50 *
  51 * The inode mark can be cleared for a number of different reasons including:
  52 * - The inode is unlinked for the last time.  (fsnotify_inode_remove)
  53 * - The inode is being evicted from cache. (fsnotify_inode_delete)
  54 * - The fs the inode is on is unmounted.  (fsnotify_inode_delete/fsnotify_unmount_inodes)
  55 * - Something explicitly requests that it be removed.  (fsnotify_destroy_mark)
  56 * - The fsnotify_group associated with the mark is going away and all such marks
  57 *   need to be cleaned up. (fsnotify_clear_marks_by_group)
  58 *
  59 * This has the very interesting property of being able to run concurrently with
  60 * any (or all) other directions.
  61 */
  62
  63#include <linux/fs.h>
  64#include <linux/init.h>
  65#include <linux/kernel.h>
  66#include <linux/kthread.h>
  67#include <linux/module.h>
  68#include <linux/mutex.h>
  69#include <linux/slab.h>
  70#include <linux/spinlock.h>
  71#include <linux/srcu.h>
  72#include <linux/ratelimit.h>
  73
  74#include <linux/atomic.h>
  75
  76#include <linux/fsnotify_backend.h>
  77#include "fsnotify.h"
  78
  79#define FSNOTIFY_REAPER_DELAY	(1)	/* 1 jiffy */
  80
  81struct srcu_struct fsnotify_mark_srcu;
  82struct kmem_cache *fsnotify_mark_connector_cachep;
  83
  84static DEFINE_SPINLOCK(destroy_lock);
  85static LIST_HEAD(destroy_list);
  86static struct fsnotify_mark_connector *connector_destroy_list;
  87
  88static void fsnotify_mark_destroy_workfn(struct work_struct *work);
  89static DECLARE_DELAYED_WORK(reaper_work, fsnotify_mark_destroy_workfn);
  90
  91static void fsnotify_connector_destroy_workfn(struct work_struct *work);
  92static DECLARE_WORK(connector_reaper_work, fsnotify_connector_destroy_workfn);
  93
  94void fsnotify_get_mark(struct fsnotify_mark *mark)
  95{
  96	WARN_ON_ONCE(!refcount_read(&mark->refcnt));
  97	refcount_inc(&mark->refcnt);
  98}
  99
 100static fsnotify_connp_t *fsnotify_object_connp(void *obj,
 101				enum fsnotify_obj_type obj_type)
 102{
 103	switch (obj_type) {
 104	case FSNOTIFY_OBJ_TYPE_INODE:
 105		return &((struct inode *)obj)->i_fsnotify_marks;
 106	case FSNOTIFY_OBJ_TYPE_VFSMOUNT:
 107		return &real_mount(obj)->mnt_fsnotify_marks;
 108	case FSNOTIFY_OBJ_TYPE_SB:
 109		return fsnotify_sb_marks(obj);
 110	case FSNOTIFY_OBJ_TYPE_MNTNS:
 111		return &((struct mnt_namespace *)obj)->n_fsnotify_marks;
 112	default:
 113		return NULL;
 114	}
 115}
 116
 117static __u32 *fsnotify_conn_mask_p(struct fsnotify_mark_connector *conn)
 118{
 119	if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
 120		return &fsnotify_conn_inode(conn)->i_fsnotify_mask;
 121	else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT)
 122		return &fsnotify_conn_mount(conn)->mnt_fsnotify_mask;
 123	else if (conn->type == FSNOTIFY_OBJ_TYPE_SB)
 124		return &fsnotify_conn_sb(conn)->s_fsnotify_mask;
 125	else if (conn->type == FSNOTIFY_OBJ_TYPE_MNTNS)
 126		return &fsnotify_conn_mntns(conn)->n_fsnotify_mask;
 127	return NULL;
 128}
 129
 130__u32 fsnotify_conn_mask(struct fsnotify_mark_connector *conn)
 131{
 132	if (WARN_ON(!fsnotify_valid_obj_type(conn->type)))
 133		return 0;
 134
 135	return READ_ONCE(*fsnotify_conn_mask_p(conn));
 136}
 137
 138static void fsnotify_get_sb_watched_objects(struct super_block *sb)
 139{
 140	atomic_long_inc(fsnotify_sb_watched_objects(sb));
 141}
 142
 143static void fsnotify_put_sb_watched_objects(struct super_block *sb)
 144{
 145	atomic_long_t *watched_objects = fsnotify_sb_watched_objects(sb);
 146
 147	/* the superblock can go away after this decrement */
 148	if (atomic_long_dec_and_test(watched_objects))
 149		wake_up_var(watched_objects);
 150}
 151
 152static void fsnotify_get_inode_ref(struct inode *inode)
 153{
 154	ihold(inode);
 155	fsnotify_get_sb_watched_objects(inode->i_sb);
 156}
 157
 158static void fsnotify_put_inode_ref(struct inode *inode)
 159{
 160	/* read ->i_sb before the inode can go away */
 161	struct super_block *sb = inode->i_sb;
 162
 163	iput(inode);
 164	fsnotify_put_sb_watched_objects(sb);
 165}
 166
 167/*
 168 * Grab or drop watched objects reference depending on whether the connector
 169 * is attached and has any marks attached.
 170 */
 171static void fsnotify_update_sb_watchers(struct super_block *sb,
 172					struct fsnotify_mark_connector *conn)
 173{
 174	struct fsnotify_sb_info *sbinfo = fsnotify_sb_info(sb);
 175	bool is_watched = conn->flags & FSNOTIFY_CONN_FLAG_IS_WATCHED;
 176	struct fsnotify_mark *first_mark = NULL;
 177	unsigned int highest_prio = 0;
 178
 179	if (conn->obj)
 180		first_mark = hlist_entry_safe(conn->list.first,
 181					      struct fsnotify_mark, obj_list);
 182	if (first_mark)
 183		highest_prio = first_mark->group->priority;
 184	if (WARN_ON(highest_prio >= __FSNOTIFY_PRIO_NUM))
 185		highest_prio = 0;
 186
 187	/*
 188	 * If the highest priority of group watching this object is prio,
 189	 * then watched object has a reference on counters [0..prio].
 190	 * Update priority >= 1 watched objects counters.
 191	 */
 192	for (unsigned int p = conn->prio + 1; p <= highest_prio; p++)
 193		atomic_long_inc(&sbinfo->watched_objects[p]);
 194	for (unsigned int p = conn->prio; p > highest_prio; p--)
 195		atomic_long_dec(&sbinfo->watched_objects[p]);
 196	conn->prio = highest_prio;
 197
 198	/* Update priority >= 0 (a.k.a total) watched objects counter */
 199	BUILD_BUG_ON(FSNOTIFY_PRIO_NORMAL != 0);
 200	if (first_mark && !is_watched) {
 201		conn->flags |= FSNOTIFY_CONN_FLAG_IS_WATCHED;
 202		fsnotify_get_sb_watched_objects(sb);
 203	} else if (!first_mark && is_watched) {
 204		conn->flags &= ~FSNOTIFY_CONN_FLAG_IS_WATCHED;
 205		fsnotify_put_sb_watched_objects(sb);
 206	}
 207}
 208
 209/*
 210 * Grab or drop inode reference for the connector if needed.
 211 *
 212 * When it's time to drop the reference, we only clear the HAS_IREF flag and
 213 * return the inode object. fsnotify_drop_object() will be resonsible for doing
 214 * iput() outside of spinlocks. This happens when last mark that wanted iref is
 215 * detached.
 216 */
 217static struct inode *fsnotify_update_iref(struct fsnotify_mark_connector *conn,
 218					  bool want_iref)
 219{
 220	bool has_iref = conn->flags & FSNOTIFY_CONN_FLAG_HAS_IREF;
 221	struct inode *inode = NULL;
 222
 223	if (conn->type != FSNOTIFY_OBJ_TYPE_INODE ||
 224	    want_iref == has_iref)
 225		return NULL;
 226
 227	if (want_iref) {
 228		/* Pin inode if any mark wants inode refcount held */
 229		fsnotify_get_inode_ref(fsnotify_conn_inode(conn));
 230		conn->flags |= FSNOTIFY_CONN_FLAG_HAS_IREF;
 231	} else {
 232		/* Unpin inode after detach of last mark that wanted iref */
 233		inode = fsnotify_conn_inode(conn);
 234		conn->flags &= ~FSNOTIFY_CONN_FLAG_HAS_IREF;
 235	}
 236
 237	return inode;
 238}
 239
 240static void *__fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
 241{
 242	u32 new_mask = 0;
 243	bool want_iref = false;
 244	struct fsnotify_mark *mark;
 245
 246	assert_spin_locked(&conn->lock);
 247	/* We can get detached connector here when inode is getting unlinked. */
 248	if (!fsnotify_valid_obj_type(conn->type))
 249		return NULL;
 250	hlist_for_each_entry(mark, &conn->list, obj_list) {
 251		if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED))
 252			continue;
 253		new_mask |= fsnotify_calc_mask(mark);
 254		if (conn->type == FSNOTIFY_OBJ_TYPE_INODE &&
 255		    !(mark->flags & FSNOTIFY_MARK_FLAG_NO_IREF))
 256			want_iref = true;
 257	}
 258	/*
 259	 * We use WRITE_ONCE() to prevent silly compiler optimizations from
 260	 * confusing readers not holding conn->lock with partial updates.
 261	 */
 262	WRITE_ONCE(*fsnotify_conn_mask_p(conn), new_mask);
 263
 264	return fsnotify_update_iref(conn, want_iref);
 265}
 266
 267static bool fsnotify_conn_watches_children(
 268					struct fsnotify_mark_connector *conn)
 269{
 270	if (conn->type != FSNOTIFY_OBJ_TYPE_INODE)
 271		return false;
 272
 273	return fsnotify_inode_watches_children(fsnotify_conn_inode(conn));
 274}
 275
 276static void fsnotify_conn_set_children_dentry_flags(
 277					struct fsnotify_mark_connector *conn)
 278{
 279	if (conn->type != FSNOTIFY_OBJ_TYPE_INODE)
 280		return;
 281
 282	fsnotify_set_children_dentry_flags(fsnotify_conn_inode(conn));
 283}
 284
 285/*
 286 * Calculate mask of events for a list of marks. The caller must make sure
 287 * connector and connector->obj cannot disappear under us.  Callers achieve
 288 * this by holding a mark->lock or mark->group->mark_mutex for a mark on this
 289 * list.
 290 */
 291void fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
 292{
 293	bool update_children;
 294
 295	if (!conn)
 296		return;
 297
 298	spin_lock(&conn->lock);
 299	update_children = !fsnotify_conn_watches_children(conn);
 300	__fsnotify_recalc_mask(conn);
 301	update_children &= fsnotify_conn_watches_children(conn);
 302	spin_unlock(&conn->lock);
 303	/*
 304	 * Set children's PARENT_WATCHED flags only if parent started watching.
 305	 * When parent stops watching, we clear false positive PARENT_WATCHED
 306	 * flags lazily in __fsnotify_parent().
 307	 */
 308	if (update_children)
 309		fsnotify_conn_set_children_dentry_flags(conn);
 310}
 311
 312/* Free all connectors queued for freeing once SRCU period ends */
 313static void fsnotify_connector_destroy_workfn(struct work_struct *work)
 314{
 315	struct fsnotify_mark_connector *conn, *free;
 316
 317	spin_lock(&destroy_lock);
 318	conn = connector_destroy_list;
 319	connector_destroy_list = NULL;
 320	spin_unlock(&destroy_lock);
 321
 322	synchronize_srcu(&fsnotify_mark_srcu);
 323	while (conn) {
 324		free = conn;
 325		conn = conn->destroy_next;
 326		kmem_cache_free(fsnotify_mark_connector_cachep, free);
 327	}
 328}
 329
 330static void *fsnotify_detach_connector_from_object(
 331					struct fsnotify_mark_connector *conn,
 332					unsigned int *type)
 333{
 334	fsnotify_connp_t *connp = fsnotify_object_connp(conn->obj, conn->type);
 335	struct super_block *sb = fsnotify_connector_sb(conn);
 336	struct inode *inode = NULL;
 337
 338	*type = conn->type;
 339	if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED)
 340		return NULL;
 341
 342	if (conn->type == FSNOTIFY_OBJ_TYPE_INODE) {
 343		inode = fsnotify_conn_inode(conn);
 344		inode->i_fsnotify_mask = 0;
 345
 346		/* Unpin inode when detaching from connector */
 347		if (!(conn->flags & FSNOTIFY_CONN_FLAG_HAS_IREF))
 348			inode = NULL;
 349	} else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT) {
 350		fsnotify_conn_mount(conn)->mnt_fsnotify_mask = 0;
 351	} else if (conn->type == FSNOTIFY_OBJ_TYPE_SB) {
 352		fsnotify_conn_sb(conn)->s_fsnotify_mask = 0;
 353	} else if (conn->type == FSNOTIFY_OBJ_TYPE_MNTNS) {
 354		fsnotify_conn_mntns(conn)->n_fsnotify_mask = 0;
 355	}
 356
 357	rcu_assign_pointer(*connp, NULL);
 358	conn->obj = NULL;
 359	conn->type = FSNOTIFY_OBJ_TYPE_DETACHED;
 360	if (sb)
 361		fsnotify_update_sb_watchers(sb, conn);
 362
 363	return inode;
 364}
 365
 366static void fsnotify_final_mark_destroy(struct fsnotify_mark *mark)
 367{
 368	struct fsnotify_group *group = mark->group;
 369
 370	if (WARN_ON_ONCE(!group))
 371		return;
 372	group->ops->free_mark(mark);
 373	fsnotify_put_group(group);
 374}
 375
 376/* Drop object reference originally held by a connector */
 377static void fsnotify_drop_object(unsigned int type, void *objp)
 378{
 379	if (!objp)
 380		return;
 381	/* Currently only inode references are passed to be dropped */
 382	if (WARN_ON_ONCE(type != FSNOTIFY_OBJ_TYPE_INODE))
 383		return;
 384	fsnotify_put_inode_ref(objp);
 385}
 386
 387void fsnotify_put_mark(struct fsnotify_mark *mark)
 388{
 389	struct fsnotify_mark_connector *conn = READ_ONCE(mark->connector);
 390	void *objp = NULL;
 391	unsigned int type = FSNOTIFY_OBJ_TYPE_DETACHED;
 392	bool free_conn = false;
 393
 394	/* Catch marks that were actually never attached to object */
 395	if (!conn) {
 396		if (refcount_dec_and_test(&mark->refcnt))
 397			fsnotify_final_mark_destroy(mark);
 398		return;
 399	}
 400
 401	/*
 402	 * We have to be careful so that traversals of obj_list under lock can
 403	 * safely grab mark reference.
 404	 */
 405	if (!refcount_dec_and_lock(&mark->refcnt, &conn->lock))
 406		return;
 407
 408	hlist_del_init_rcu(&mark->obj_list);
 409	if (hlist_empty(&conn->list)) {
 410		objp = fsnotify_detach_connector_from_object(conn, &type);
 411		free_conn = true;
 412	} else {
 413		struct super_block *sb = fsnotify_connector_sb(conn);
 414
 415		/* Update watched objects after detaching mark */
 416		if (sb)
 417			fsnotify_update_sb_watchers(sb, conn);
 418		objp = __fsnotify_recalc_mask(conn);
 419		type = conn->type;
 420	}
 421	WRITE_ONCE(mark->connector, NULL);
 422	spin_unlock(&conn->lock);
 423
 424	fsnotify_drop_object(type, objp);
 425
 426	if (free_conn) {
 427		spin_lock(&destroy_lock);
 428		conn->destroy_next = connector_destroy_list;
 429		connector_destroy_list = conn;
 430		spin_unlock(&destroy_lock);
 431		queue_work(system_dfl_wq, &connector_reaper_work);
 432	}
 433	/*
 434	 * Note that we didn't update flags telling whether inode cares about
 435	 * what's happening with children. We update these flags from
 436	 * __fsnotify_parent() lazily when next event happens on one of our
 437	 * children.
 438	 */
 439	spin_lock(&destroy_lock);
 440	list_add(&mark->g_list, &destroy_list);
 441	spin_unlock(&destroy_lock);
 442	queue_delayed_work(system_dfl_wq, &reaper_work,
 443			   FSNOTIFY_REAPER_DELAY);
 444}
 445EXPORT_SYMBOL_GPL(fsnotify_put_mark);
 446
 447/*
 448 * Get mark reference when we found the mark via lockless traversal of object
 449 * list. Mark can be already removed from the list by now and on its way to be
 450 * destroyed once SRCU period ends.
 451 *
 452 * Also pin the group so it doesn't disappear under us.
 453 */
 454static bool fsnotify_get_mark_safe(struct fsnotify_mark *mark)
 455{
 456	if (!mark)
 457		return true;
 458
 459	if (refcount_inc_not_zero(&mark->refcnt)) {
 460		spin_lock(&mark->lock);
 461		if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) {
 462			/* mark is attached, group is still alive then */
 463			atomic_inc(&mark->group->user_waits);
 464			spin_unlock(&mark->lock);
 465			return true;
 466		}
 467		spin_unlock(&mark->lock);
 468		fsnotify_put_mark(mark);
 469	}
 470	return false;
 471}
 472
 473/*
 474 * Puts marks and wakes up group destruction if necessary.
 475 *
 476 * Pairs with fsnotify_get_mark_safe()
 477 */
 478static void fsnotify_put_mark_wake(struct fsnotify_mark *mark)
 479{
 480	if (mark) {
 481		struct fsnotify_group *group = mark->group;
 482
 483		fsnotify_put_mark(mark);
 484		/*
 485		 * We abuse notification_waitq on group shutdown for waiting for
 486		 * all marks pinned when waiting for userspace.
 487		 */
 488		if (atomic_dec_and_test(&group->user_waits) && group->shutdown)
 489			wake_up(&group->notification_waitq);
 490	}
 491}
 492
 493bool fsnotify_prepare_user_wait(struct fsnotify_iter_info *iter_info)
 494	__releases(&fsnotify_mark_srcu)
 495{
 496	int type;
 497
 498	fsnotify_foreach_iter_type(type) {
 499		/* This can fail if mark is being removed */
 500		if (!fsnotify_get_mark_safe(iter_info->marks[type])) {
 501			__release(&fsnotify_mark_srcu);
 502			goto fail;
 503		}
 504	}
 505
 506	/*
 507	 * Now that both marks are pinned by refcount in the inode / vfsmount
 508	 * lists, we can drop SRCU lock, and safely resume the list iteration
 509	 * once userspace returns.
 510	 */
 511	srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
 512
 513	return true;
 514
 515fail:
 516	for (type--; type >= 0; type--)
 517		fsnotify_put_mark_wake(iter_info->marks[type]);
 518	return false;
 519}
 520
 521void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
 522	__acquires(&fsnotify_mark_srcu)
 523{
 524	int type;
 525
 526	iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
 527	fsnotify_foreach_iter_type(type)
 528		fsnotify_put_mark_wake(iter_info->marks[type]);
 529}
 530
 531/*
 532 * Mark mark as detached, remove it from group list. Mark still stays in object
 533 * list until its last reference is dropped. Note that we rely on mark being
 534 * removed from group list before corresponding reference to it is dropped. In
 535 * particular we rely on mark->connector being valid while we hold
 536 * group->mark_mutex if we found the mark through g_list.
 537 *
 538 * Must be called with group->mark_mutex held. The caller must either hold
 539 * reference to the mark or be protected by fsnotify_mark_srcu.
 540 */
 541void fsnotify_detach_mark(struct fsnotify_mark *mark)
 542{
 543	fsnotify_group_assert_locked(mark->group);
 544	WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
 545		     refcount_read(&mark->refcnt) < 1 +
 546			!!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
 547
 548	spin_lock(&mark->lock);
 549	/* something else already called this function on this mark */
 550	if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 551		spin_unlock(&mark->lock);
 552		return;
 553	}
 554	mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
 555	list_del_init(&mark->g_list);
 556	spin_unlock(&mark->lock);
 557
 558	/* Drop mark reference acquired in fsnotify_add_mark_locked() */
 559	fsnotify_put_mark(mark);
 560}
 561
 562/*
 563 * Free fsnotify mark. The mark is actually only marked as being freed.  The
 564 * freeing is actually happening only once last reference to the mark is
 565 * dropped from a workqueue which first waits for srcu period end.
 566 *
 567 * Caller must have a reference to the mark or be protected by
 568 * fsnotify_mark_srcu.
 569 */
 570void fsnotify_free_mark(struct fsnotify_mark *mark)
 571{
 572	struct fsnotify_group *group = mark->group;
 573
 574	spin_lock(&mark->lock);
 575	/* something else already called this function on this mark */
 576	if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
 577		spin_unlock(&mark->lock);
 578		return;
 579	}
 580	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
 581	spin_unlock(&mark->lock);
 582
 583	/*
 584	 * Some groups like to know that marks are being freed.  This is a
 585	 * callback to the group function to let it know that this mark
 586	 * is being freed.
 587	 */
 588	if (group->ops->freeing_mark)
 589		group->ops->freeing_mark(mark, group);
 590}
 591
 592void fsnotify_destroy_mark(struct fsnotify_mark *mark,
 593			   struct fsnotify_group *group)
 594{
 595	fsnotify_group_lock(group);
 596	fsnotify_detach_mark(mark);
 597	fsnotify_group_unlock(group);
 598	fsnotify_free_mark(mark);
 599}
 600EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
 601
 602/*
 603 * Sorting function for lists of fsnotify marks.
 604 *
 605 * Fanotify supports different notification classes (reflected as priority of
 606 * notification group). Events shall be passed to notification groups in
 607 * decreasing priority order. To achieve this marks in notification lists for
 608 * inodes and vfsmounts are sorted so that priorities of corresponding groups
 609 * are descending.
 610 *
 611 * Furthermore correct handling of the ignore mask requires processing inode
 612 * and vfsmount marks of each group together. Using the group address as
 613 * further sort criterion provides a unique sorting order and thus we can
 614 * merge inode and vfsmount lists of marks in linear time and find groups
 615 * present in both lists.
 616 *
 617 * A return value of 1 signifies that b has priority over a.
 618 * A return value of 0 signifies that the two marks have to be handled together.
 619 * A return value of -1 signifies that a has priority over b.
 620 */
 621int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
 622{
 623	if (a == b)
 624		return 0;
 625	if (!a)
 626		return 1;
 627	if (!b)
 628		return -1;
 629	if (a->priority < b->priority)
 630		return 1;
 631	if (a->priority > b->priority)
 632		return -1;
 633	if (a < b)
 634		return 1;
 635	return -1;
 636}
 637
 638static int fsnotify_attach_info_to_sb(struct super_block *sb)
 639{
 640	struct fsnotify_sb_info *sbinfo;
 641
 642	/* sb info is freed on fsnotify_sb_delete() */
 643	sbinfo = kzalloc(sizeof(*sbinfo), GFP_KERNEL);
 644	if (!sbinfo)
 645		return -ENOMEM;
 646
 647	/*
 648	 * cmpxchg() provides the barrier so that callers of fsnotify_sb_info()
 649	 * will observe an initialized structure
 650	 */
 651	if (cmpxchg(&sb->s_fsnotify_info, NULL, sbinfo)) {
 652		/* Someone else created sbinfo for us */
 653		kfree(sbinfo);
 654	}
 655	return 0;
 656}
 657
 658static int fsnotify_attach_connector_to_object(fsnotify_connp_t *connp,
 659					       void *obj, unsigned int obj_type)
 660{
 661	struct fsnotify_mark_connector *conn;
 662
 663	conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
 664	if (!conn)
 665		return -ENOMEM;
 666	spin_lock_init(&conn->lock);
 667	INIT_HLIST_HEAD(&conn->list);
 668	conn->flags = 0;
 669	conn->prio = 0;
 670	conn->type = obj_type;
 671	conn->obj = obj;
 672
 673	/*
 674	 * cmpxchg() provides the barrier so that readers of *connp can see
 675	 * only initialized structure
 676	 */
 677	if (cmpxchg(connp, NULL, conn)) {
 678		/* Someone else created list structure for us */
 679		kmem_cache_free(fsnotify_mark_connector_cachep, conn);
 680	}
 681	return 0;
 682}
 683
 684/*
 685 * Get mark connector, make sure it is alive and return with its lock held.
 686 * This is for users that get connector pointer from inode or mount. Users that
 687 * hold reference to a mark on the list may directly lock connector->lock as
 688 * they are sure list cannot go away under them.
 689 */
 690static struct fsnotify_mark_connector *fsnotify_grab_connector(
 691						fsnotify_connp_t *connp)
 692{
 693	struct fsnotify_mark_connector *conn;
 694	int idx;
 695
 696	idx = srcu_read_lock(&fsnotify_mark_srcu);
 697	conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
 698	if (!conn)
 699		goto out;
 700	spin_lock(&conn->lock);
 701	if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED) {
 702		spin_unlock(&conn->lock);
 703		srcu_read_unlock(&fsnotify_mark_srcu, idx);
 704		return NULL;
 705	}
 706out:
 707	srcu_read_unlock(&fsnotify_mark_srcu, idx);
 708	return conn;
 709}
 710
 711/*
 712 * Add mark into proper place in given list of marks. These marks may be used
 713 * for the fsnotify backend to determine which event types should be delivered
 714 * to which group and for which inodes. These marks are ordered according to
 715 * priority, highest number first, and then by the group's location in memory.
 716 */
 717static int fsnotify_add_mark_list(struct fsnotify_mark *mark, void *obj,
 718				  unsigned int obj_type, int add_flags)
 719{
 720	struct super_block *sb = fsnotify_object_sb(obj, obj_type);
 721	struct fsnotify_mark *lmark, *last = NULL;
 722	struct fsnotify_mark_connector *conn;
 723	fsnotify_connp_t *connp;
 724	int cmp;
 725	int err = 0;
 726
 727	if (WARN_ON(!fsnotify_valid_obj_type(obj_type)))
 728		return -EINVAL;
 729
 730	/*
 731	 * Attach the sb info before attaching a connector to any object on sb.
 732	 * The sb info will remain attached as long as sb lives.
 733	 */
 734	if (sb && !fsnotify_sb_info(sb)) {
 735		err = fsnotify_attach_info_to_sb(sb);
 736		if (err)
 737			return err;
 738	}
 739
 740	connp = fsnotify_object_connp(obj, obj_type);
 741restart:
 742	spin_lock(&mark->lock);
 743	conn = fsnotify_grab_connector(connp);
 744	if (!conn) {
 745		spin_unlock(&mark->lock);
 746		err = fsnotify_attach_connector_to_object(connp, obj, obj_type);
 747		if (err)
 748			return err;
 749		goto restart;
 750	}
 751
 752	/* is mark the first mark? */
 753	if (hlist_empty(&conn->list)) {
 754		hlist_add_head_rcu(&mark->obj_list, &conn->list);
 755		goto added;
 756	}
 757
 758	/* should mark be in the middle of the current list? */
 759	hlist_for_each_entry(lmark, &conn->list, obj_list) {
 760		last = lmark;
 761
 762		if ((lmark->group == mark->group) &&
 763		    (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
 764		    !(mark->group->flags & FSNOTIFY_GROUP_DUPS)) {
 765			err = -EEXIST;
 766			goto out_err;
 767		}
 768
 769		cmp = fsnotify_compare_groups(lmark->group, mark->group);
 770		if (cmp >= 0) {
 771			hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
 772			goto added;
 773		}
 774	}
 775
 776	BUG_ON(last == NULL);
 777	/* mark should be the last entry.  last is the current last entry */
 778	hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
 779added:
 780	if (sb)
 781		fsnotify_update_sb_watchers(sb, conn);
 782	/*
 783	 * Since connector is attached to object using cmpxchg() we are
 784	 * guaranteed that connector initialization is fully visible by anyone
 785	 * seeing mark->connector set.
 786	 */
 787	WRITE_ONCE(mark->connector, conn);
 788out_err:
 789	spin_unlock(&conn->lock);
 790	spin_unlock(&mark->lock);
 791	return err;
 792}
 793
 794/*
 795 * Attach an initialized mark to a given group and fs object.
 796 * These marks may be used for the fsnotify backend to determine which
 797 * event types should be delivered to which group.
 798 */
 799int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
 800			     void *obj, unsigned int obj_type,
 801			     int add_flags)
 802{
 803	struct fsnotify_group *group = mark->group;
 804	int ret = 0;
 805
 806	fsnotify_group_assert_locked(group);
 807
 808	/*
 809	 * LOCKING ORDER!!!!
 810	 * group->mark_mutex
 811	 * mark->lock
 812	 * mark->connector->lock
 813	 */
 814	spin_lock(&mark->lock);
 815	mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
 816
 817	list_add(&mark->g_list, &group->marks_list);
 818	fsnotify_get_mark(mark); /* for g_list */
 819	spin_unlock(&mark->lock);
 820
 821	ret = fsnotify_add_mark_list(mark, obj, obj_type, add_flags);
 822	if (ret)
 823		goto err;
 824
 825	fsnotify_recalc_mask(mark->connector);
 826
 827	return ret;
 828err:
 829	spin_lock(&mark->lock);
 830	mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
 831			 FSNOTIFY_MARK_FLAG_ATTACHED);
 832	list_del_init(&mark->g_list);
 833	spin_unlock(&mark->lock);
 834
 835	fsnotify_put_mark(mark);
 836	return ret;
 837}
 838
 839int fsnotify_add_mark(struct fsnotify_mark *mark, void *obj,
 840		      unsigned int obj_type, int add_flags)
 841{
 842	int ret;
 843	struct fsnotify_group *group = mark->group;
 844
 845	fsnotify_group_lock(group);
 846	ret = fsnotify_add_mark_locked(mark, obj, obj_type, add_flags);
 847	fsnotify_group_unlock(group);
 848	return ret;
 849}
 850EXPORT_SYMBOL_GPL(fsnotify_add_mark);
 851
 852/*
 853 * Given a list of marks, find the mark associated with given group. If found
 854 * take a reference to that mark and return it, else return NULL.
 855 */
 856struct fsnotify_mark *fsnotify_find_mark(void *obj, unsigned int obj_type,
 857					 struct fsnotify_group *group)
 858{
 859	fsnotify_connp_t *connp = fsnotify_object_connp(obj, obj_type);
 860	struct fsnotify_mark_connector *conn;
 861	struct fsnotify_mark *mark;
 862
 863	if (!connp)
 864		return NULL;
 865
 866	conn = fsnotify_grab_connector(connp);
 867	if (!conn)
 868		return NULL;
 869
 870	hlist_for_each_entry(mark, &conn->list, obj_list) {
 871		if (mark->group == group &&
 872		    (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 873			fsnotify_get_mark(mark);
 874			spin_unlock(&conn->lock);
 875			return mark;
 876		}
 877	}
 878	spin_unlock(&conn->lock);
 879	return NULL;
 880}
 881EXPORT_SYMBOL_GPL(fsnotify_find_mark);
 882
 883/* Clear any marks in a group with given type mask */
 884void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
 885				   unsigned int obj_type)
 886{
 887	struct fsnotify_mark *lmark, *mark;
 888	LIST_HEAD(to_free);
 889	struct list_head *head = &to_free;
 890
 891	/* Skip selection step if we want to clear all marks. */
 892	if (obj_type == FSNOTIFY_OBJ_TYPE_ANY) {
 893		head = &group->marks_list;
 894		goto clear;
 895	}
 896	/*
 897	 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
 898	 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
 899	 * to_free list so we have to use mark_mutex even when accessing that
 900	 * list. And freeing mark requires us to drop mark_mutex. So we can
 901	 * reliably free only the first mark in the list. That's why we first
 902	 * move marks to free to to_free list in one go and then free marks in
 903	 * to_free list one by one.
 904	 */
 905	fsnotify_group_lock(group);
 906	list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
 907		if (mark->connector->type == obj_type)
 908			list_move(&mark->g_list, &to_free);
 909	}
 910	fsnotify_group_unlock(group);
 911
 912clear:
 913	while (1) {
 914		fsnotify_group_lock(group);
 915		if (list_empty(head)) {
 916			fsnotify_group_unlock(group);
 917			break;
 918		}
 919		mark = list_first_entry(head, struct fsnotify_mark, g_list);
 920		fsnotify_get_mark(mark);
 921		fsnotify_detach_mark(mark);
 922		fsnotify_group_unlock(group);
 923		fsnotify_free_mark(mark);
 924		fsnotify_put_mark(mark);
 925	}
 926}
 927
 928/* Destroy all marks attached to an object via connector */
 929void fsnotify_destroy_marks(fsnotify_connp_t *connp)
 930{
 931	struct fsnotify_mark_connector *conn;
 932	struct fsnotify_mark *mark, *old_mark = NULL;
 933	void *objp;
 934	unsigned int type;
 935
 936	conn = fsnotify_grab_connector(connp);
 937	if (!conn)
 938		return;
 939	/*
 940	 * We have to be careful since we can race with e.g.
 941	 * fsnotify_clear_marks_by_group() and once we drop the conn->lock, the
 942	 * list can get modified. However we are holding mark reference and
 943	 * thus our mark cannot be removed from obj_list so we can continue
 944	 * iteration after regaining conn->lock.
 945	 */
 946	hlist_for_each_entry(mark, &conn->list, obj_list) {
 947		fsnotify_get_mark(mark);
 948		spin_unlock(&conn->lock);
 949		if (old_mark)
 950			fsnotify_put_mark(old_mark);
 951		old_mark = mark;
 952		fsnotify_destroy_mark(mark, mark->group);
 953		spin_lock(&conn->lock);
 954	}
 955	/*
 956	 * Detach list from object now so that we don't pin inode until all
 957	 * mark references get dropped. It would lead to strange results such
 958	 * as delaying inode deletion or blocking unmount.
 959	 */
 960	objp = fsnotify_detach_connector_from_object(conn, &type);
 961	spin_unlock(&conn->lock);
 962	if (old_mark)
 963		fsnotify_put_mark(old_mark);
 964	fsnotify_drop_object(type, objp);
 965}
 966
 967/*
 968 * Nothing fancy, just initialize lists and locks and counters.
 969 */
 970void fsnotify_init_mark(struct fsnotify_mark *mark,
 971			struct fsnotify_group *group)
 972{
 973	memset(mark, 0, sizeof(*mark));
 974	spin_lock_init(&mark->lock);
 975	refcount_set(&mark->refcnt, 1);
 976	fsnotify_get_group(group);
 977	mark->group = group;
 978	WRITE_ONCE(mark->connector, NULL);
 979}
 980EXPORT_SYMBOL_GPL(fsnotify_init_mark);
 981
 982/*
 983 * Destroy all marks in destroy_list, waits for SRCU period to finish before
 984 * actually freeing marks.
 985 */
 986static void fsnotify_mark_destroy_workfn(struct work_struct *work)
 987{
 988	struct fsnotify_mark *mark, *next;
 989	struct list_head private_destroy_list;
 990
 991	spin_lock(&destroy_lock);
 992	/* exchange the list head */
 993	list_replace_init(&destroy_list, &private_destroy_list);
 994	spin_unlock(&destroy_lock);
 995
 996	synchronize_srcu(&fsnotify_mark_srcu);
 997
 998	list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
 999		list_del_init(&mark->g_list);
1000		fsnotify_final_mark_destroy(mark);
1001	}
1002}
1003
1004/* Wait for all marks queued for destruction to be actually destroyed */
1005void fsnotify_wait_marks_destroyed(void)
1006{
1007	flush_delayed_work(&reaper_work);
1008}
1009EXPORT_SYMBOL_GPL(fsnotify_wait_marks_destroyed);