blob: 79334ed5daa7fee4802c5f0a30bb58dde88dbbdc [file] [log] [blame]
Pekka Enbergcbf8f0f2005-11-07 01:01:09 -08001RCU-based dcache locking model
2==============================
3
4On many workloads, the most common operation on dcache is to look up a
5dentry, given a parent dentry and the name of the child. Typically,
6for every open(), stat() etc., the dentry corresponding to the
7pathname will be looked up by walking the tree starting with the first
8component of the pathname and using that dentry along with the next
9component to look up the next level and so on. Since it is a frequent
10operation for workloads like multiuser environments and web servers,
11it is important to optimize this path.
12
13Prior to 2.5.10, dcache_lock was acquired in d_lookup and thus in
14every component during path look-up. Since 2.5.10 onwards, fast-walk
15algorithm changed this by holding the dcache_lock at the beginning and
16walking as many cached path component dentries as possible. This
17significantly decreases the number of acquisition of
18dcache_lock. However it also increases the lock hold time
19significantly and affects performance in large SMP machines. Since
202.5.62 kernel, dcache has been using a new locking model that uses RCU
21to make dcache look-up lock-free.
22
23The current dcache locking model is not very different from the
24existing dcache locking model. Prior to 2.5.62 kernel, dcache_lock
25protected the hash chain, d_child, d_alias, d_lru lists as well as
26d_inode and several other things like mount look-up. RCU-based changes
27affect only the way the hash chain is protected. For everything else
28the dcache_lock must be taken for both traversing as well as
29updating. The hash chain updates too take the dcache_lock. The
30significant change is the way d_lookup traverses the hash chain, it
31doesn't acquire the dcache_lock for this and rely on RCU to ensure
32that the dentry has not been *freed*.
33
34
35Dcache locking details
36======================
37
38For many multi-user workloads, open() and stat() on files are very
39frequently occurring operations. Both involve walking of path names to
40find the dentry corresponding to the concerned file. In 2.4 kernel,
41dcache_lock was held during look-up of each path component. Contention
42and cache-line bouncing of this global lock caused significant
43scalability problems. With the introduction of RCU in Linux kernel,
44this was worked around by making the look-up of path components during
45path walking lock-free.
46
47
48Safe lock-free look-up of dcache hash table
49===========================================
50
51Dcache is a complex data structure with the hash table entries also
52linked together in other lists. In 2.4 kernel, dcache_lock protected
53all the lists. We applied RCU only on hash chain walking. The rest of
54the lists are still protected by dcache_lock. Some of the important
55changes are :
56
571. The deletion from hash chain is done using hlist_del_rcu() macro
58 which doesn't initialize next pointer of the deleted dentry and
59 this allows us to walk safely lock-free while a deletion is
60 happening.
61
622. Insertion of a dentry into the hash table is done using
63 hlist_add_head_rcu() which take care of ordering the writes - the
64 writes to the dentry must be visible before the dentry is
Paul E. McKenney4c540052010-01-14 16:10:57 -080065 inserted. This works in conjunction with hlist_for_each_rcu(),
66 which has since been replaced by hlist_for_each_entry_rcu(), while
Pekka Enbergcbf8f0f2005-11-07 01:01:09 -080067 walking the hash chain. The only requirement is that all
68 initialization to the dentry must be done before
69 hlist_add_head_rcu() since we don't have dcache_lock protection
70 while traversing the hash chain. This isn't different from the
71 existing code.
72
733. The dentry looked up without holding dcache_lock by cannot be
74 returned for walking if it is unhashed. It then may have a NULL
75 d_inode or other bogosity since RCU doesn't protect the other
76 fields in the dentry. We therefore use a flag DCACHE_UNHASHED to
77 indicate unhashed dentries and use this in conjunction with a
78 per-dentry lock (d_lock). Once looked up without the dcache_lock,
79 we acquire the per-dentry lock (d_lock) and check if the dentry is
80 unhashed. If so, the look-up is failed. If not, the reference count
81 of the dentry is increased and the dentry is returned.
82
834. Once a dentry is looked up, it must be ensured during the path walk
84 for that component it doesn't go away. In pre-2.5.10 code, this was
85 done holding a reference to the dentry. dcache_rcu does the same.
86 In some sense, dcache_rcu path walking looks like the pre-2.5.10
87 version.
88
895. All dentry hash chain updates must take the dcache_lock as well as
90 the per-dentry lock in that order. dput() does this to ensure that
91 a dentry that has just been looked up in another CPU doesn't get
92 deleted before dget() can be done on it.
93
946. There are several ways to do reference counting of RCU protected
95 objects. One such example is in ipv4 route cache where deferred
96 freeing (using call_rcu()) is done as soon as the reference count
97 goes to zero. This cannot be done in the case of dentries because
98 tearing down of dentries require blocking (dentry_iput()) which
99 isn't supported from RCU callbacks. Instead, tearing down of
100 dentries happen synchronously in dput(), but actual freeing happens
101 later when RCU grace period is over. This allows safe lock-free
102 walking of the hash chains, but a matched dentry may have been
103 partially torn down. The checking of DCACHE_UNHASHED flag with
104 d_lock held detects such dentries and prevents them from being
105 returned from look-up.
106
107
108Maintaining POSIX rename semantics
109==================================
110
111Since look-up of dentries is lock-free, it can race against a
112concurrent rename operation. For example, during rename of file A to
113B, look-up of either A or B must succeed. So, if look-up of B happens
114after A has been removed from the hash chain but not added to the new
115hash chain, it may fail. Also, a comparison while the name is being
116written concurrently by a rename may result in false positive matches
117violating rename semantics. Issues related to race with rename are
118handled as described below :
119
1201. Look-up can be done in two ways - d_lookup() which is safe from
121 simultaneous renames and __d_lookup() which is not. If
122 __d_lookup() fails, it must be followed up by a d_lookup() to
123 correctly determine whether a dentry is in the hash table or
124 not. d_lookup() protects look-ups using a sequence lock
125 (rename_lock).
126
1272. The name associated with a dentry (d_name) may be changed if a
128 rename is allowed to happen simultaneously. To avoid memcmp() in
129 __d_lookup() go out of bounds due to a rename and false positive
130 comparison, the name comparison is done while holding the
131 per-dentry lock. This prevents concurrent renames during this
132 operation.
133
1343. Hash table walking during look-up may move to a different bucket as
135 the current dentry is moved to a different bucket due to rename.
136 But we use hlists in dcache hash table and they are
137 null-terminated. So, even if a dentry moves to a different bucket,
138 hash chain walk will terminate. [with a list_head list, it may not
139 since termination is when the list_head in the original bucket is
140 reached]. Since we redo the d_parent check and compare name while
141 holding d_lock, lock-free look-up will not race against d_move().
142
1434. There can be a theoretical race when a dentry keeps coming back to
144 original bucket due to double moves. Due to this look-up may
145 consider that it has never moved and can end up in a infinite loop.
146 But this is not any worse that theoretical livelocks we already
147 have in the kernel.
148
149
150Important guidelines for filesystem developers related to dcache_rcu
151====================================================================
152
1531. Existing dcache interfaces (pre-2.5.62) exported to filesystem
154 don't change. Only dcache internal implementation changes. However
155 filesystems *must not* delete from the dentry hash chains directly
156 using the list macros like allowed earlier. They must use dcache
157 APIs like d_drop() or __d_drop() depending on the situation.
158
1592. d_flags is now protected by a per-dentry lock (d_lock). All access
160 to d_flags must be protected by it.
161
1623. For a hashed dentry, checking of d_count needs to be protected by
163 d_lock.
164
165
166Papers and other documentation on dcache locking
167================================================
168
1691. Scaling dcache with RCU (http://linuxjournal.com/article.php?sid=7124).
170
1712. http://lse.sourceforge.net/locking/dcache/dcache.html
172
173
174