Based on kernel version 4.8. Page generated on 2016-10-06 23:18 EST.
1 Started by: Ingo Molnar <firstname.lastname@example.org> 2 3 Background 4 ---------- 5 6 what are robust futexes? To answer that, we first need to understand 7 what futexes are: normal futexes are special types of locks that in the 8 noncontended case can be acquired/released from userspace without having 9 to enter the kernel. 10 11 A futex is in essence a user-space address, e.g. a 32-bit lock variable 12 field. If userspace notices contention (the lock is already owned and 13 someone else wants to grab it too) then the lock is marked with a value 14 that says "there's a waiter pending", and the sys_futex(FUTEX_WAIT) 15 syscall is used to wait for the other guy to release it. The kernel 16 creates a 'futex queue' internally, so that it can later on match up the 17 waiter with the waker - without them having to know about each other. 18 When the owner thread releases the futex, it notices (via the variable 19 value) that there were waiter(s) pending, and does the 20 sys_futex(FUTEX_WAKE) syscall to wake them up. Once all waiters have 21 taken and released the lock, the futex is again back to 'uncontended' 22 state, and there's no in-kernel state associated with it. The kernel 23 completely forgets that there ever was a futex at that address. This 24 method makes futexes very lightweight and scalable. 25 26 "Robustness" is about dealing with crashes while holding a lock: if a 27 process exits prematurely while holding a pthread_mutex_t lock that is 28 also shared with some other process (e.g. yum segfaults while holding a 29 pthread_mutex_t, or yum is kill -9-ed), then waiters for that lock need 30 to be notified that the last owner of the lock exited in some irregular 31 way. 32 33 To solve such types of problems, "robust mutex" userspace APIs were 34 created: pthread_mutex_lock() returns an error value if the owner exits 35 prematurely - and the new owner can decide whether the data protected by 36 the lock can be recovered safely. 37 38 There is a big conceptual problem with futex based mutexes though: it is 39 the kernel that destroys the owner task (e.g. due to a SEGFAULT), but 40 the 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) 42 then the kernel has no information to clean up after the held lock! 43 Userspace has no chance to clean up after the lock either - userspace is 44 the one that crashes, so it has no opportunity to clean up. Catch-22. 45 46 In practice, when e.g. yum is kill -9-ed (or segfaults), a system reboot 47 is needed to release that futex based lock. This is one of the leading 48 bugreports against yum. 49 50 To solve this problem, the traditional approach was to extend the vma 51 (virtual memory area descriptor) concept to have a notion of 'pending 52 robust futexes attached to this area'. This approach requires 3 new 53 syscall variants to sys_futex(): FUTEX_REGISTER, FUTEX_DEREGISTER and 54 FUTEX_RECOVER. At do_exit() time, all vmas are searched to see whether 55 they have a robust_head set. This approach has two fundamental problems 56 left: 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 64 The second disadvantage is a real killer: pthread_exit() takes around 1 65 microsecond on Linux, but with thousands (or tens of thousands) of vmas 66 every pthread_exit() takes a millisecond or more, also totally 67 destroying the CPU's L1 and L2 caches! 68 69 This is very much noticeable even for normal process sys_exit_group() 70 calls: the kernel has to do the vma scanning unconditionally! (this is 71 because the kernel has no knowledge about how many robust futexes there 72 are to be cleaned up, because a robust futex might have been registered 73 in another task, and the futex variable might have been simply mmap()-ed 74 into this process's address space). 75 76 This huge overhead forced the creation of CONFIG_FUTEX_ROBUST so that 77 normal kernels can turn it off, but worse than that: the overhead makes 78 robust futexes impractical for any type of generic Linux distribution. 79 80 So something had to be done. 81 82 New approach to robust futexes 83 ------------------------------ 84 85 At the heart of this new approach there is a per-thread private list of 86 robust locks that userspace is holding (maintained by glibc) - which 87 userspace list is registered with the kernel via a new syscall [this 88 registration happens at most once per thread lifetime]. At do_exit() 89 time, the kernel checks this user-space list: are there any robust futex 90 locks to be cleaned up? 91 92 In the common case, at do_exit() time, there is no list registered, so 93 the cost of robust futexes is just a simple current->robust_list != NULL 94 comparison. If the thread has registered a list, then normally the list 95 is empty. If the thread/process crashed or terminated in some incorrect 96 way then the list might be non-empty: in this case the kernel carefully 97 walks the list [not trusting it], and marks all locks that are owned by 98 this thread with the FUTEX_OWNER_DIED bit, and wakes up one waiter (if 99 any). 100 101 The list is guaranteed to be private and per-thread at do_exit() time, 102 so it can be accessed by the kernel in a lockless way. 103 104 There is one race possible though: since adding to and removing from the 105 list is done after the futex is acquired by glibc, there is a few 106 instructions window for the thread (or process) to die there, leaving 107 the futex hung. To protect against this possibility, userspace (glibc) 108 also maintains a simple per-thread 'list_op_pending' field, to allow the 109 kernel to clean up if the thread dies after acquiring the lock, but just 110 before it could have added itself to the list. Glibc sets this 111 list_op_pending field before it tries to acquire the futex, and clears 112 it after the list-add (or list-remove) has finished. 113 114 That's all that is needed - all the rest of robust-futex cleanup is done 115 in userspace [just like with the previous patches]. 116 117 Ulrich Drepper has implemented the necessary glibc support for this new 118 mechanism, which fully enables robust mutexes. 119 120 Key differences of this userspace-list based approach, compared to the 121 vma 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 don't 130 need any extra per-lock syscalls. Robust mutexes thus become a very 131 lightweight primitive - so they don't 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. 143 144 Performance 145 ----------- 146 147 I have benchmarked the time needed for the kernel to process a list of 1 148 million (!) 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 152 153 I have also measured an approach where glibc does the lock notification 154 [which it currently does for !pshared robust mutexes], and that took 256 155 msecs - clearly slower, due to the 1 million FUTEX_WAKE syscalls 156 userspace had to do. 157 158 (1 million held locks are unheard of - we expect at most a handful of 159 locks to be held at a time. Nevertheless it's nice to know that this 160 approach scales nicely.) 161 162 Implementation details 163 ---------------------- 164 165 The patch adds two new syscalls: one to register the userspace list, and 166 one 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); 175 176 List registration is very fast: the pointer is simply stored in 177 current->robust_list. [Note that in the future, if robust futexes become 178 widespread, we could extend sys_clone() to register a robust-list head 179 for new threads, without the need of another syscall.] 180 181 So there is virtually zero overhead for tasks not using robust futexes, 182 and even for robust futex users, there is only one extra syscall per 183 thread lifetime, and the cleanup operation, if it happens, is fast and 184 straightforward. The kernel doesn't have any internal distinction between 185 robust and normal futexes. 186 187 If a futex is found to be held at exit time, the kernel sets the 188 following bit of the futex word: 189 190 #define FUTEX_OWNER_DIED 0x40000000 191 192 and wakes up the next futex waiter (if any). User-space does the rest of 193 the cleanup. 194 195 Otherwise, robust futexes are acquired by glibc by putting the TID into 196 the futex field atomically. Waiters set the FUTEX_WAITERS bit: 197 198 #define FUTEX_WAITERS 0x80000000 199 200 and the remaining bits are for the TID. 201 202 Testing, architecture support 203 ----------------------------- 204 205 I've tested the new syscalls on x86 and x86_64, and have made sure the 206 parsing of the userspace list is robust [ ;-) ] even if the list is 207 deliberately corrupted. 208 209 i386 and x86_64 syscalls are wired up at the moment, and Ulrich has 210 tested the new glibc code (on x86_64 and i386), and it works for his 211 robust-mutex testcases. 212 213 All other architectures should build just fine too - but they won't have 214 the new syscalls yet. 215 216 Architectures need to implement the new futex_atomic_cmpxchg_inatomic() 217 inline function before writing up the syscalls (that function returns 218 -ENOSYS right now).