robust-futexes.txt 9.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218
  1. Started by: Ingo Molnar <mingo@redhat.com>
  2. Background
  3. ----------
  4. what are robust futexes? To answer that, we first need to understand
  5. what futexes are: normal futexes are special types of locks that in the
  6. noncontended case can be acquired/released from userspace without having
  7. to enter the kernel.
  8. A futex is in essence a user-space address, e.g. a 32-bit lock variable
  9. field. If userspace notices contention (the lock is already owned and
  10. someone else wants to grab it too) then the lock is marked with a value
  11. that says "there's a waiter pending", and the sys_futex(FUTEX_WAIT)
  12. syscall is used to wait for the other guy to release it. The kernel
  13. creates a 'futex queue' internally, so that it can later on match up the
  14. waiter with the waker - without them having to know about each other.
  15. When the owner thread releases the futex, it notices (via the variable
  16. value) that there were waiter(s) pending, and does the
  17. sys_futex(FUTEX_WAKE) syscall to wake them up. Once all waiters have
  18. taken and released the lock, the futex is again back to 'uncontended'
  19. state, and there's no in-kernel state associated with it. The kernel
  20. completely forgets that there ever was a futex at that address. This
  21. method makes futexes very lightweight and scalable.
  22. "Robustness" is about dealing with crashes while holding a lock: if a
  23. process exits prematurely while holding a pthread_mutex_t lock that is
  24. also shared with some other process (e.g. yum segfaults while holding a
  25. pthread_mutex_t, or yum is kill -9-ed), then waiters for that lock need
  26. to be notified that the last owner of the lock exited in some irregular
  27. way.
  28. To solve such types of problems, "robust mutex" userspace APIs were
  29. created: pthread_mutex_lock() returns an error value if the owner exits
  30. prematurely - and the new owner can decide whether the data protected by
  31. the lock can be recovered safely.
  32. There is a big conceptual problem with futex based mutexes though: it is
  33. the kernel that destroys the owner task (e.g. due to a SEGFAULT), but
  34. the kernel cannot help with the cleanup: if there is no 'futex queue'
  35. (and in most cases there is none, futexes being fast lightweight locks)
  36. then the kernel has no information to clean up after the held lock!
  37. Userspace has no chance to clean up after the lock either - userspace is
  38. the one that crashes, so it has no opportunity to clean up. Catch-22.
  39. In practice, when e.g. yum is kill -9-ed (or segfaults), a system reboot
  40. is needed to release that futex based lock. This is one of the leading
  41. bugreports against yum.
  42. To solve this problem, the traditional approach was to extend the vma
  43. (virtual memory area descriptor) concept to have a notion of 'pending
  44. robust futexes attached to this area'. This approach requires 3 new
  45. syscall variants to sys_futex(): FUTEX_REGISTER, FUTEX_DEREGISTER and
  46. FUTEX_RECOVER. At do_exit() time, all vmas are searched to see whether
  47. they have a robust_head set. This approach has two fundamental problems
  48. left:
  49. - it has quite complex locking and race scenarios. The vma-based
  50. approach had been pending for years, but they are still not completely
  51. reliable.
  52. - they have to scan _every_ vma at sys_exit() time, per thread!
  53. The second disadvantage is a real killer: pthread_exit() takes around 1
  54. microsecond on Linux, but with thousands (or tens of thousands) of vmas
  55. every pthread_exit() takes a millisecond or more, also totally
  56. destroying the CPU's L1 and L2 caches!
  57. This is very much noticeable even for normal process sys_exit_group()
  58. calls: the kernel has to do the vma scanning unconditionally! (this is
  59. because the kernel has no knowledge about how many robust futexes there
  60. are to be cleaned up, because a robust futex might have been registered
  61. in another task, and the futex variable might have been simply mmap()-ed
  62. into this process's address space).
  63. This huge overhead forced the creation of CONFIG_FUTEX_ROBUST so that
  64. normal kernels can turn it off, but worse than that: the overhead makes
  65. robust futexes impractical for any type of generic Linux distribution.
  66. So something had to be done.
  67. New approach to robust futexes
  68. ------------------------------
  69. At the heart of this new approach there is a per-thread private list of
  70. robust locks that userspace is holding (maintained by glibc) - which
  71. userspace list is registered with the kernel via a new syscall [this
  72. registration happens at most once per thread lifetime]. At do_exit()
  73. time, the kernel checks this user-space list: are there any robust futex
  74. locks to be cleaned up?
  75. In the common case, at do_exit() time, there is no list registered, so
  76. the cost of robust futexes is just a simple current->robust_list != NULL
  77. comparison. If the thread has registered a list, then normally the list
  78. is empty. If the thread/process crashed or terminated in some incorrect
  79. way then the list might be non-empty: in this case the kernel carefully
  80. walks the list [not trusting it], and marks all locks that are owned by
  81. this thread with the FUTEX_OWNER_DIED bit, and wakes up one waiter (if
  82. any).
  83. The list is guaranteed to be private and per-thread at do_exit() time,
  84. so it can be accessed by the kernel in a lockless way.
  85. There is one race possible though: since adding to and removing from the
  86. list is done after the futex is acquired by glibc, there is a few
  87. instructions window for the thread (or process) to die there, leaving
  88. the futex hung. To protect against this possibility, userspace (glibc)
  89. also maintains a simple per-thread 'list_op_pending' field, to allow the
  90. kernel to clean up if the thread dies after acquiring the lock, but just
  91. before it could have added itself to the list. Glibc sets this
  92. list_op_pending field before it tries to acquire the futex, and clears
  93. it after the list-add (or list-remove) has finished.
  94. That's all that is needed - all the rest of robust-futex cleanup is done
  95. in userspace [just like with the previous patches].
  96. Ulrich Drepper has implemented the necessary glibc support for this new
  97. mechanism, which fully enables robust mutexes.
  98. Key differences of this userspace-list based approach, compared to the
  99. vma based method:
  100. - it's much, much faster: at thread exit time, there's no need to loop
  101. over every vma (!), which the VM-based method has to do. Only a very
  102. simple 'is the list empty' op is done.
  103. - no VM changes are needed - 'struct address_space' is left alone.
  104. - no registration of individual locks is needed: robust mutexes dont
  105. need any extra per-lock syscalls. Robust mutexes thus become a very
  106. lightweight primitive - so they dont force the application designer
  107. to do a hard choice between performance and robustness - robust
  108. mutexes are just as fast.
  109. - no per-lock kernel allocation happens.
  110. - no resource limits are needed.
  111. - no kernel-space recovery call (FUTEX_RECOVER) is needed.
  112. - the implementation and the locking is "obvious", and there are no
  113. interactions with the VM.
  114. Performance
  115. -----------
  116. I have benchmarked the time needed for the kernel to process a list of 1
  117. million (!) held locks, using the new method [on a 2GHz CPU]:
  118. - with FUTEX_WAIT set [contended mutex]: 130 msecs
  119. - without FUTEX_WAIT set [uncontended mutex]: 30 msecs
  120. I have also measured an approach where glibc does the lock notification
  121. [which it currently does for !pshared robust mutexes], and that took 256
  122. msecs - clearly slower, due to the 1 million FUTEX_WAKE syscalls
  123. userspace had to do.
  124. (1 million held locks are unheard of - we expect at most a handful of
  125. locks to be held at a time. Nevertheless it's nice to know that this
  126. approach scales nicely.)
  127. Implementation details
  128. ----------------------
  129. The patch adds two new syscalls: one to register the userspace list, and
  130. one to query the registered list pointer:
  131. asmlinkage long
  132. sys_set_robust_list(struct robust_list_head __user *head,
  133. size_t len);
  134. asmlinkage long
  135. sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
  136. size_t __user *len_ptr);
  137. List registration is very fast: the pointer is simply stored in
  138. current->robust_list. [Note that in the future, if robust futexes become
  139. widespread, we could extend sys_clone() to register a robust-list head
  140. for new threads, without the need of another syscall.]
  141. So there is virtually zero overhead for tasks not using robust futexes,
  142. and even for robust futex users, there is only one extra syscall per
  143. thread lifetime, and the cleanup operation, if it happens, is fast and
  144. straightforward. The kernel doesn't have any internal distinction between
  145. robust and normal futexes.
  146. If a futex is found to be held at exit time, the kernel sets the
  147. following bit of the futex word:
  148. #define FUTEX_OWNER_DIED 0x40000000
  149. and wakes up the next futex waiter (if any). User-space does the rest of
  150. the cleanup.
  151. Otherwise, robust futexes are acquired by glibc by putting the TID into
  152. the futex field atomically. Waiters set the FUTEX_WAITERS bit:
  153. #define FUTEX_WAITERS 0x80000000
  154. and the remaining bits are for the TID.
  155. Testing, architecture support
  156. -----------------------------
  157. i've tested the new syscalls on x86 and x86_64, and have made sure the
  158. parsing of the userspace list is robust [ ;-) ] even if the list is
  159. deliberately corrupted.
  160. i386 and x86_64 syscalls are wired up at the moment, and Ulrich has
  161. tested the new glibc code (on x86_64 and i386), and it works for his
  162. robust-mutex testcases.
  163. All other architectures should build just fine too - but they wont have
  164. the new syscalls yet.
  165. Architectures need to implement the new futex_atomic_cmpxchg_inatomic()
  166. inline function before writing up the syscalls (that function returns
  167. -ENOSYS right now).