1Started by: Ingo Molnar <mingo@redhat.com>
2
3Background
4----------
5
6what are robust futexes? To answer that, we first need to understand
7what futexes are: normal futexes are special types of locks that in the
8noncontended case can be acquired/released from userspace without having
9to enter the kernel.
10
11A futex is in essence a user-space address, e.g. a 32-bit lock variable
12field. If userspace notices contention (the lock is already owned and
13someone else wants to grab it too) then the lock is marked with a value
14that says "there's a waiter pending", and the sys_futex(FUTEX_WAIT)
15syscall is used to wait for the other guy to release it. The kernel
16creates a 'futex queue' internally, so that it can later on match up the
17waiter with the waker - without them having to know about each other.
18When the owner thread releases the futex, it notices (via the variable
19value) that there were waiter(s) pending, and does the
20sys_futex(FUTEX_WAKE) syscall to wake them up.  Once all waiters have
21taken and released the lock, the futex is again back to 'uncontended'
22state, and there's no in-kernel state associated with it. The kernel
23completely forgets that there ever was a futex at that address. This
24method makes futexes very lightweight and scalable.
25
26"Robustness" is about dealing with crashes while holding a lock: if a
27process exits prematurely while holding a pthread_mutex_t lock that is
28also shared with some other process (e.g. yum segfaults while holding a
29pthread_mutex_t, or yum is kill -9-ed), then waiters for that lock need
30to be notified that the last owner of the lock exited in some irregular
31way.
32
33To solve such types of problems, "robust mutex" userspace APIs were
34created: pthread_mutex_lock() returns an error value if the owner exits
35prematurely - and the new owner can decide whether the data protected by
36the lock can be recovered safely.
37
38There is a big conceptual problem with futex based mutexes though: it is
39the kernel that destroys the owner task (e.g. due to a SEGFAULT), but
40the kernel cannot help with the cleanup: if there is no 'futex queue'
41(and in most cases there is none, futexes being fast lightweight locks)
42then the kernel has no information to clean up after the held lock!
43Userspace has no chance to clean up after the lock either - userspace is
44the one that crashes, so it has no opportunity to clean up. Catch-22.
45
46In practice, when e.g. yum is kill -9-ed (or segfaults), a system reboot
47is needed to release that futex based lock. This is one of the leading
48bugreports against yum.
49
50To solve this problem, the traditional approach was to extend the vma
51(virtual memory area descriptor) concept to have a notion of 'pending
52robust futexes attached to this area'. This approach requires 3 new
53syscall variants to sys_futex(): FUTEX_REGISTER, FUTEX_DEREGISTER and
54FUTEX_RECOVER. At do_exit() time, all vmas are searched to see whether
55they have a robust_head set. This approach has two fundamental problems
56left:
57
58 - it has quite complex locking and race scenarios. The vma-based
59   approach had been pending for years, but they are still not completely
60   reliable.
61
62 - they have to scan _every_ vma at sys_exit() time, per thread!
63
64The second disadvantage is a real killer: pthread_exit() takes around 1
65microsecond on Linux, but with thousands (or tens of thousands) of vmas
66every pthread_exit() takes a millisecond or more, also totally
67destroying the CPU's L1 and L2 caches!
68
69This is very much noticeable even for normal process sys_exit_group()
70calls: the kernel has to do the vma scanning unconditionally! (this is
71because the kernel has no knowledge about how many robust futexes there
72are to be cleaned up, because a robust futex might have been registered
73in another task, and the futex variable might have been simply mmap()-ed
74into this process's address space).
75
76This huge overhead forced the creation of CONFIG_FUTEX_ROBUST so that
77normal kernels can turn it off, but worse than that: the overhead makes
78robust futexes impractical for any type of generic Linux distribution.
79
80So something had to be done.
81
82New approach to robust futexes
83------------------------------
84
85At the heart of this new approach there is a per-thread private list of
86robust locks that userspace is holding (maintained by glibc) - which
87userspace list is registered with the kernel via a new syscall [this
88registration happens at most once per thread lifetime]. At do_exit()
89time, the kernel checks this user-space list: are there any robust futex
90locks to be cleaned up?
91
92In the common case, at do_exit() time, there is no list registered, so
93the cost of robust futexes is just a simple current->robust_list != NULL
94comparison. If the thread has registered a list, then normally the list
95is empty. If the thread/process crashed or terminated in some incorrect
96way then the list might be non-empty: in this case the kernel carefully
97walks the list [not trusting it], and marks all locks that are owned by
98this thread with the FUTEX_OWNER_DIED bit, and wakes up one waiter (if
99any).
100
101The list is guaranteed to be private and per-thread at do_exit() time,
102so it can be accessed by the kernel in a lockless way.
103
104There is one race possible though: since adding to and removing from the
105list is done after the futex is acquired by glibc, there is a few
106instructions window for the thread (or process) to die there, leaving
107the futex hung. To protect against this possibility, userspace (glibc)
108also maintains a simple per-thread 'list_op_pending' field, to allow the
109kernel to clean up if the thread dies after acquiring the lock, but just
110before it could have added itself to the list. Glibc sets this
111list_op_pending field before it tries to acquire the futex, and clears
112it after the list-add (or list-remove) has finished.
113
114That's all that is needed - all the rest of robust-futex cleanup is done
115in userspace [just like with the previous patches].
116
117Ulrich Drepper has implemented the necessary glibc support for this new
118mechanism, which fully enables robust mutexes.
119
120Key differences of this userspace-list based approach, compared to the
121vma based method:
122
123 - it's much, much faster: at thread exit time, there's no need to loop
124   over every vma (!), which the VM-based method has to do. Only a very
125   simple 'is the list empty' op is done.
126
127 - no VM changes are needed - 'struct address_space' is left alone.
128
129 - no registration of individual locks is needed: robust mutexes dont
130   need any extra per-lock syscalls. Robust mutexes thus become a very
131   lightweight primitive - so they dont force the application designer
132   to do a hard choice between performance and robustness - robust
133   mutexes are just as fast.
134
135 - no per-lock kernel allocation happens.
136
137 - no resource limits are needed.
138
139 - no kernel-space recovery call (FUTEX_RECOVER) is needed.
140
141 - the implementation and the locking is "obvious", and there are no
142   interactions with the VM.
143
144Performance
145-----------
146
147I have benchmarked the time needed for the kernel to process a list of 1
148million (!) held locks, using the new method [on a 2GHz CPU]:
149
150 - with FUTEX_WAIT set [contended mutex]: 130 msecs
151 - without FUTEX_WAIT set [uncontended mutex]: 30 msecs
152
153I have also measured an approach where glibc does the lock notification
154[which it currently does for !pshared robust mutexes], and that took 256
155msecs - clearly slower, due to the 1 million FUTEX_WAKE syscalls
156userspace had to do.
157
158(1 million held locks are unheard of - we expect at most a handful of
159locks to be held at a time. Nevertheless it's nice to know that this
160approach scales nicely.)
161
162Implementation details
163----------------------
164
165The patch adds two new syscalls: one to register the userspace list, and
166one to query the registered list pointer:
167
168 asmlinkage long
169 sys_set_robust_list(struct robust_list_head __user *head,
170                     size_t len);
171
172 asmlinkage long
173 sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
174                     size_t __user *len_ptr);
175
176List registration is very fast: the pointer is simply stored in
177current->robust_list. [Note that in the future, if robust futexes become
178widespread, we could extend sys_clone() to register a robust-list head
179for new threads, without the need of another syscall.]
180
181So there is virtually zero overhead for tasks not using robust futexes,
182and even for robust futex users, there is only one extra syscall per
183thread lifetime, and the cleanup operation, if it happens, is fast and
184straightforward. The kernel doesn't have any internal distinction between
185robust and normal futexes.
186
187If a futex is found to be held at exit time, the kernel sets the
188following bit of the futex word:
189
190	#define FUTEX_OWNER_DIED        0x40000000
191
192and wakes up the next futex waiter (if any). User-space does the rest of
193the cleanup.
194
195Otherwise, robust futexes are acquired by glibc by putting the TID into
196the futex field atomically. Waiters set the FUTEX_WAITERS bit:
197
198	#define FUTEX_WAITERS           0x80000000
199
200and the remaining bits are for the TID.
201
202Testing, architecture support
203-----------------------------
204
205i've tested the new syscalls on x86 and x86_64, and have made sure the
206parsing of the userspace list is robust [ ;-) ] even if the list is
207deliberately corrupted.
208
209i386 and x86_64 syscalls are wired up at the moment, and Ulrich has
210tested the new glibc code (on x86_64 and i386), and it works for his
211robust-mutex testcases.
212
213All other architectures should build just fine too - but they won't have
214the new syscalls yet.
215
216Architectures need to implement the new futex_atomic_cmpxchg_inatomic()
217inline function before writing up the syscalls (that function returns
218-ENOSYS right now).
219