Back to home page

LXR

 
 

    


0001 Started by Paul Jackson <pj@sgi.com>
0002 
0003 The robust futex ABI
0004 --------------------
0005 
0006 Robust_futexes provide a mechanism that is used in addition to normal
0007 futexes, for kernel assist of cleanup of held locks on task exit.
0008 
0009 The interesting data as to what futexes a thread is holding is kept on a
0010 linked list in user space, where it can be updated efficiently as locks
0011 are taken and dropped, without kernel intervention.  The only additional
0012 kernel intervention required for robust_futexes above and beyond what is
0013 required for futexes is:
0014 
0015  1) a one time call, per thread, to tell the kernel where its list of
0016     held robust_futexes begins, and
0017  2) internal kernel code at exit, to handle any listed locks held
0018     by the exiting thread.
0019 
0020 The existing normal futexes already provide a "Fast Userspace Locking"
0021 mechanism, which handles uncontested locking without needing a system
0022 call, and handles contested locking by maintaining a list of waiting
0023 threads in the kernel.  Options on the sys_futex(2) system call support
0024 waiting on a particular futex, and waking up the next waiter on a
0025 particular futex.
0026 
0027 For robust_futexes to work, the user code (typically in a library such
0028 as glibc linked with the application) has to manage and place the
0029 necessary list elements exactly as the kernel expects them.  If it fails
0030 to do so, then improperly listed locks will not be cleaned up on exit,
0031 probably causing deadlock or other such failure of the other threads
0032 waiting on the same locks.
0033 
0034 A thread that anticipates possibly using robust_futexes should first
0035 issue the system call:
0036 
0037     asmlinkage long
0038     sys_set_robust_list(struct robust_list_head __user *head, size_t len);
0039 
0040 The pointer 'head' points to a structure in the threads address space
0041 consisting of three words.  Each word is 32 bits on 32 bit arch's, or 64
0042 bits on 64 bit arch's, and local byte order.  Each thread should have
0043 its own thread private 'head'.
0044 
0045 If a thread is running in 32 bit compatibility mode on a 64 native arch
0046 kernel, then it can actually have two such structures - one using 32 bit
0047 words for 32 bit compatibility mode, and one using 64 bit words for 64
0048 bit native mode.  The kernel, if it is a 64 bit kernel supporting 32 bit
0049 compatibility mode, will attempt to process both lists on each task
0050 exit, if the corresponding sys_set_robust_list() call has been made to
0051 setup that list.
0052 
0053   The first word in the memory structure at 'head' contains a
0054   pointer to a single linked list of 'lock entries', one per lock,
0055   as described below.  If the list is empty, the pointer will point
0056   to itself, 'head'.  The last 'lock entry' points back to the 'head'.
0057 
0058   The second word, called 'offset', specifies the offset from the
0059   address of the associated 'lock entry', plus or minus, of what will
0060   be called the 'lock word', from that 'lock entry'.  The 'lock word'
0061   is always a 32 bit word, unlike the other words above.  The 'lock
0062   word' holds 3 flag bits in the upper 3 bits, and the thread id (TID)
0063   of the thread holding the lock in the bottom 29 bits.  See further
0064   below for a description of the flag bits.
0065 
0066   The third word, called 'list_op_pending', contains transient copy of
0067   the address of the 'lock entry', during list insertion and removal,
0068   and is needed to correctly resolve races should a thread exit while
0069   in the middle of a locking or unlocking operation.
0070 
0071 Each 'lock entry' on the single linked list starting at 'head' consists
0072 of just a single word, pointing to the next 'lock entry', or back to
0073 'head' if there are no more entries.  In addition, nearby to each 'lock
0074 entry', at an offset from the 'lock entry' specified by the 'offset'
0075 word, is one 'lock word'.
0076 
0077 The 'lock word' is always 32 bits, and is intended to be the same 32 bit
0078 lock variable used by the futex mechanism, in conjunction with
0079 robust_futexes.  The kernel will only be able to wakeup the next thread
0080 waiting for a lock on a threads exit if that next thread used the futex
0081 mechanism to register the address of that 'lock word' with the kernel.
0082 
0083 For each futex lock currently held by a thread, if it wants this
0084 robust_futex support for exit cleanup of that lock, it should have one
0085 'lock entry' on this list, with its associated 'lock word' at the
0086 specified 'offset'.  Should a thread die while holding any such locks,
0087 the kernel will walk this list, mark any such locks with a bit
0088 indicating their holder died, and wakeup the next thread waiting for
0089 that lock using the futex mechanism.
0090 
0091 When a thread has invoked the above system call to indicate it
0092 anticipates using robust_futexes, the kernel stores the passed in 'head'
0093 pointer for that task.  The task may retrieve that value later on by
0094 using the system call:
0095 
0096     asmlinkage long
0097     sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
0098                         size_t __user *len_ptr);
0099 
0100 It is anticipated that threads will use robust_futexes embedded in
0101 larger, user level locking structures, one per lock.  The kernel
0102 robust_futex mechanism doesn't care what else is in that structure, so
0103 long as the 'offset' to the 'lock word' is the same for all
0104 robust_futexes used by that thread.  The thread should link those locks
0105 it currently holds using the 'lock entry' pointers.  It may also have
0106 other links between the locks, such as the reverse side of a double
0107 linked list, but that doesn't matter to the kernel.
0108 
0109 By keeping its locks linked this way, on a list starting with a 'head'
0110 pointer known to the kernel, the kernel can provide to a thread the
0111 essential service available for robust_futexes, which is to help clean
0112 up locks held at the time of (a perhaps unexpectedly) exit.
0113 
0114 Actual locking and unlocking, during normal operations, is handled
0115 entirely by user level code in the contending threads, and by the
0116 existing futex mechanism to wait for, and wakeup, locks.  The kernels
0117 only essential involvement in robust_futexes is to remember where the
0118 list 'head' is, and to walk the list on thread exit, handling locks
0119 still held by the departing thread, as described below.
0120 
0121 There may exist thousands of futex lock structures in a threads shared
0122 memory, on various data structures, at a given point in time. Only those
0123 lock structures for locks currently held by that thread should be on
0124 that thread's robust_futex linked lock list a given time.
0125 
0126 A given futex lock structure in a user shared memory region may be held
0127 at different times by any of the threads with access to that region. The
0128 thread currently holding such a lock, if any, is marked with the threads
0129 TID in the lower 29 bits of the 'lock word'.
0130 
0131 When adding or removing a lock from its list of held locks, in order for
0132 the kernel to correctly handle lock cleanup regardless of when the task
0133 exits (perhaps it gets an unexpected signal 9 in the middle of
0134 manipulating this list), the user code must observe the following
0135 protocol on 'lock entry' insertion and removal:
0136 
0137 On insertion:
0138  1) set the 'list_op_pending' word to the address of the 'lock entry'
0139     to be inserted,
0140  2) acquire the futex lock,
0141  3) add the lock entry, with its thread id (TID) in the bottom 29 bits
0142     of the 'lock word', to the linked list starting at 'head', and
0143  4) clear the 'list_op_pending' word.
0144 
0145 On removal:
0146  1) set the 'list_op_pending' word to the address of the 'lock entry'
0147     to be removed,
0148  2) remove the lock entry for this lock from the 'head' list,
0149  3) release the futex lock, and
0150  4) clear the 'lock_op_pending' word.
0151 
0152 On exit, the kernel will consider the address stored in
0153 'list_op_pending' and the address of each 'lock word' found by walking
0154 the list starting at 'head'.  For each such address, if the bottom 29
0155 bits of the 'lock word' at offset 'offset' from that address equals the
0156 exiting threads TID, then the kernel will do two things:
0157 
0158  1) if bit 31 (0x80000000) is set in that word, then attempt a futex
0159     wakeup on that address, which will waken the next thread that has
0160     used to the futex mechanism to wait on that address, and
0161  2) atomically set  bit 30 (0x40000000) in the 'lock word'.
0162 
0163 In the above, bit 31 was set by futex waiters on that lock to indicate
0164 they were waiting, and bit 30 is set by the kernel to indicate that the
0165 lock owner died holding the lock.
0166 
0167 The kernel exit code will silently stop scanning the list further if at
0168 any point:
0169 
0170  1) the 'head' pointer or an subsequent linked list pointer
0171     is not a valid address of a user space word
0172  2) the calculated location of the 'lock word' (address plus
0173     'offset') is not the valid address of a 32 bit user space
0174     word
0175  3) if the list contains more than 1 million (subject to
0176     future kernel configuration changes) elements.
0177 
0178 When the kernel sees a list entry whose 'lock word' doesn't have the
0179 current threads TID in the lower 29 bits, it does nothing with that
0180 entry, and goes on to the next entry.
0181 
0182 Bit 29 (0x20000000) of the 'lock word' is reserved for future use.