[PATCH] lightweight robust futexes: docs

Add robust-futex documentation.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This commit is contained in:
Ingo Molnar 2006-03-27 01:16:23 -08:00 коммит произвёл Linus Torvalds
Родитель 0771dfefc9
Коммит 2eec9ad91f
2 изменённых файлов: 402 добавлений и 0 удалений

Просмотреть файл

@ -0,0 +1,184 @@
Started by Paul Jackson <pj@sgi.com>
The robust futex ABI
--------------------
Robust_futexes provide a mechanism that is used in addition to normal
futexes, for kernel assist of cleanup of held locks on task exit.
The interesting data as to what futexes a thread is holding is kept on a
linked list in user space, where it can be updated efficiently as locks
are taken and dropped, without kernel intervention. The only additional
kernel intervention required for robust_futexes above and beyond what is
required for futexes is:
1) a one time call, per thread, to tell the kernel where its list of
held robust_futexes begins, and
2) internal kernel code at exit, to handle any listed locks held
by the exiting thread.
The existing normal futexes already provide a "Fast Userspace Locking"
mechanism, which handles uncontested locking without needing a system
call, and handles contested locking by maintaining a list of waiting
threads in the kernel. Options on the sys_futex(2) system call support
waiting on a particular futex, and waking up the next waiter on a
particular futex.
For robust_futexes to work, the user code (typically in a library such
as glibc linked with the application) has to manage and place the
necessary list elements exactly as the kernel expects them. If it fails
to do so, then improperly listed locks will not be cleaned up on exit,
probably causing deadlock or other such failure of the other threads
waiting on the same locks.
A thread that anticipates possibly using robust_futexes should first
issue the system call:
asmlinkage long
sys_set_robust_list(struct robust_list_head __user *head, size_t len);
The pointer 'head' points to a structure in the threads address space
consisting of three words. Each word is 32 bits on 32 bit arch's, or 64
bits on 64 bit arch's, and local byte order. Each thread should have
its own thread private 'head'.
If a thread is running in 32 bit compatibility mode on a 64 native arch
kernel, then it can actually have two such structures - one using 32 bit
words for 32 bit compatibility mode, and one using 64 bit words for 64
bit native mode. The kernel, if it is a 64 bit kernel supporting 32 bit
compatibility mode, will attempt to process both lists on each task
exit, if the corresponding sys_set_robust_list() call has been made to
setup that list.
The first word in the memory structure at 'head' contains a
pointer to a single linked list of 'lock entries', one per lock,
as described below. If the list is empty, the pointer will point
to itself, 'head'. The last 'lock entry' points back to the 'head'.
The second word, called 'offset', specifies the offset from the
address of the associated 'lock entry', plus or minus, of what will
be called the 'lock word', from that 'lock entry'. The 'lock word'
is always a 32 bit word, unlike the other words above. The 'lock
word' holds 3 flag bits in the upper 3 bits, and the thread id (TID)
of the thread holding the lock in the bottom 29 bits. See further
below for a description of the flag bits.
The third word, called 'list_op_pending', contains transient copy of
the address of the 'lock entry', during list insertion and removal,
and is needed to correctly resolve races should a thread exit while
in the middle of a locking or unlocking operation.
Each 'lock entry' on the single linked list starting at 'head' consists
of just a single word, pointing to the next 'lock entry', or back to
'head' if there are no more entries. In addition, nearby to each 'lock
entry', at an offset from the 'lock entry' specified by the 'offset'
word, is one 'lock word'.
The 'lock word' is always 32 bits, and is intended to be the same 32 bit
lock variable used by the futex mechanism, in conjunction with
robust_futexes. The kernel will only be able to wakeup the next thread
waiting for a lock on a threads exit if that next thread used the futex
mechanism to register the address of that 'lock word' with the kernel.
For each futex lock currently held by a thread, if it wants this
robust_futex support for exit cleanup of that lock, it should have one
'lock entry' on this list, with its associated 'lock word' at the
specified 'offset'. Should a thread die while holding any such locks,
the kernel will walk this list, mark any such locks with a bit
indicating their holder died, and wakeup the next thread waiting for
that lock using the futex mechanism.
When a thread has invoked the above system call to indicate it
anticipates using robust_futexes, the kernel stores the passed in 'head'
pointer for that task. The task may retrieve that value later on by
using the system call:
asmlinkage long
sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
size_t __user *len_ptr);
It is anticipated that threads will use robust_futexes embedded in
larger, user level locking structures, one per lock. The kernel
robust_futex mechanism doesn't care what else is in that structure, so
long as the 'offset' to the 'lock word' is the same for all
robust_futexes used by that thread. The thread should link those locks
it currently holds using the 'lock entry' pointers. It may also have
other links between the locks, such as the reverse side of a double
linked list, but that doesn't matter to the kernel.
By keeping its locks linked this way, on a list starting with a 'head'
pointer known to the kernel, the kernel can provide to a thread the
essential service available for robust_futexes, which is to help clean
up locks held at the time of (a perhaps unexpectedly) exit.
Actual locking and unlocking, during normal operations, is handled
entirely by user level code in the contending threads, and by the
existing futex mechanism to wait for, and wakeup, locks. The kernels
only essential involvement in robust_futexes is to remember where the
list 'head' is, and to walk the list on thread exit, handling locks
still held by the departing thread, as described below.
There may exist thousands of futex lock structures in a threads shared
memory, on various data structures, at a given point in time. Only those
lock structures for locks currently held by that thread should be on
that thread's robust_futex linked lock list a given time.
A given futex lock structure in a user shared memory region may be held
at different times by any of the threads with access to that region. The
thread currently holding such a lock, if any, is marked with the threads
TID in the lower 29 bits of the 'lock word'.
When adding or removing a lock from its list of held locks, in order for
the kernel to correctly handle lock cleanup regardless of when the task
exits (perhaps it gets an unexpected signal 9 in the middle of
manipulating this list), the user code must observe the following
protocol on 'lock entry' insertion and removal:
On insertion:
1) set the 'list_op_pending' word to the address of the 'lock word'
to be inserted,
2) acquire the futex lock,
3) add the lock entry, with its thread id (TID) in the bottom 29 bits
of the 'lock word', to the linked list starting at 'head', and
4) clear the 'list_op_pending' word.
XXX I am particularly unsure of the following -pj XXX
On removal:
1) set the 'list_op_pending' word to the address of the 'lock word'
to be removed,
2) remove the lock entry for this lock from the 'head' list,
2) release the futex lock, and
2) clear the 'lock_op_pending' word.
On exit, the kernel will consider the address stored in
'list_op_pending' and the address of each 'lock word' found by walking
the list starting at 'head'. For each such address, if the bottom 29
bits of the 'lock word' at offset 'offset' from that address equals the
exiting threads TID, then the kernel will do two things:
1) if bit 31 (0x80000000) is set in that word, then attempt a futex
wakeup on that address, which will waken the next thread that has
used to the futex mechanism to wait on that address, and
2) atomically set bit 30 (0x40000000) in the 'lock word'.
In the above, bit 31 was set by futex waiters on that lock to indicate
they were waiting, and bit 30 is set by the kernel to indicate that the
lock owner died holding the lock.
The kernel exit code will silently stop scanning the list further if at
any point:
1) the 'head' pointer or an subsequent linked list pointer
is not a valid address of a user space word
2) the calculated location of the 'lock word' (address plus
'offset') is not the valud address of a 32 bit user space
word
3) if the list contains more than 1 million (subject to
future kernel configuration changes) elements.
When the kernel sees a list entry whose 'lock word' doesn't have the
current threads TID in the lower 29 bits, it does nothing with that
entry, and goes on to the next entry.
Bit 29 (0x20000000) of the 'lock word' is reserved for future use.

Просмотреть файл

@ -0,0 +1,218 @@
Started by: Ingo Molnar <mingo@redhat.com>
Background
----------
what are robust futexes? To answer that, we first need to understand
what futexes are: normal futexes are special types of locks that in the
noncontended case can be acquired/released from userspace without having
to enter the kernel.
A futex is in essence a user-space address, e.g. a 32-bit lock variable
field. If userspace notices contention (the lock is already owned and
someone else wants to grab it too) then the lock is marked with a value
that says "there's a waiter pending", and the sys_futex(FUTEX_WAIT)
syscall is used to wait for the other guy to release it. The kernel
creates a 'futex queue' internally, so that it can later on match up the
waiter with the waker - without them having to know about each other.
When the owner thread releases the futex, it notices (via the variable
value) that there were waiter(s) pending, and does the
sys_futex(FUTEX_WAKE) syscall to wake them up. Once all waiters have
taken and released the lock, the futex is again back to 'uncontended'
state, and there's no in-kernel state associated with it. The kernel
completely forgets that there ever was a futex at that address. This
method makes futexes very lightweight and scalable.
"Robustness" is about dealing with crashes while holding a lock: if a
process exits prematurely while holding a pthread_mutex_t lock that is
also shared with some other process (e.g. yum segfaults while holding a
pthread_mutex_t, or yum is kill -9-ed), then waiters for that lock need
to be notified that the last owner of the lock exited in some irregular
way.
To solve such types of problems, "robust mutex" userspace APIs were
created: pthread_mutex_lock() returns an error value if the owner exits
prematurely - and the new owner can decide whether the data protected by
the lock can be recovered safely.
There is a big conceptual problem with futex based mutexes though: it is
the kernel that destroys the owner task (e.g. due to a SEGFAULT), but
the kernel cannot help with the cleanup: if there is no 'futex queue'
(and in most cases there is none, futexes being fast lightweight locks)
then the kernel has no information to clean up after the held lock!
Userspace has no chance to clean up after the lock either - userspace is
the one that crashes, so it has no opportunity to clean up. Catch-22.
In practice, when e.g. yum is kill -9-ed (or segfaults), a system reboot
is needed to release that futex based lock. This is one of the leading
bugreports against yum.
To solve this problem, the traditional approach was to extend the vma
(virtual memory area descriptor) concept to have a notion of 'pending
robust futexes attached to this area'. This approach requires 3 new
syscall variants to sys_futex(): FUTEX_REGISTER, FUTEX_DEREGISTER and
FUTEX_RECOVER. At do_exit() time, all vmas are searched to see whether
they have a robust_head set. This approach has two fundamental problems
left:
- it has quite complex locking and race scenarios. The vma-based
approach had been pending for years, but they are still not completely
reliable.
- they have to scan _every_ vma at sys_exit() time, per thread!
The second disadvantage is a real killer: pthread_exit() takes around 1
microsecond on Linux, but with thousands (or tens of thousands) of vmas
every pthread_exit() takes a millisecond or more, also totally
destroying the CPU's L1 and L2 caches!
This is very much noticeable even for normal process sys_exit_group()
calls: the kernel has to do the vma scanning unconditionally! (this is
because the kernel has no knowledge about how many robust futexes there
are to be cleaned up, because a robust futex might have been registered
in another task, and the futex variable might have been simply mmap()-ed
into this process's address space).
This huge overhead forced the creation of CONFIG_FUTEX_ROBUST so that
normal kernels can turn it off, but worse than that: the overhead makes
robust futexes impractical for any type of generic Linux distribution.
So something had to be done.
New approach to robust futexes
------------------------------
At the heart of this new approach there is a per-thread private list of
robust locks that userspace is holding (maintained by glibc) - which
userspace list is registered with the kernel via a new syscall [this
registration happens at most once per thread lifetime]. At do_exit()
time, the kernel checks this user-space list: are there any robust futex
locks to be cleaned up?
In the common case, at do_exit() time, there is no list registered, so
the cost of robust futexes is just a simple current->robust_list != NULL
comparison. If the thread has registered a list, then normally the list
is empty. If the thread/process crashed or terminated in some incorrect
way then the list might be non-empty: in this case the kernel carefully
walks the list [not trusting it], and marks all locks that are owned by
this thread with the FUTEX_OWNER_DEAD bit, and wakes up one waiter (if
any).
The list is guaranteed to be private and per-thread at do_exit() time,
so it can be accessed by the kernel in a lockless way.
There is one race possible though: since adding to and removing from the
list is done after the futex is acquired by glibc, there is a few
instructions window for the thread (or process) to die there, leaving
the futex hung. To protect against this possibility, userspace (glibc)
also maintains a simple per-thread 'list_op_pending' field, to allow the
kernel to clean up if the thread dies after acquiring the lock, but just
before it could have added itself to the list. Glibc sets this
list_op_pending field before it tries to acquire the futex, and clears
it after the list-add (or list-remove) has finished.
That's all that is needed - all the rest of robust-futex cleanup is done
in userspace [just like with the previous patches].
Ulrich Drepper has implemented the necessary glibc support for this new
mechanism, which fully enables robust mutexes.
Key differences of this userspace-list based approach, compared to the
vma based method:
- it's much, much faster: at thread exit time, there's no need to loop
over every vma (!), which the VM-based method has to do. Only a very
simple 'is the list empty' op is done.
- no VM changes are needed - 'struct address_space' is left alone.
- no registration of individual locks is needed: robust mutexes dont
need any extra per-lock syscalls. Robust mutexes thus become a very
lightweight primitive - so they dont force the application designer
to do a hard choice between performance and robustness - robust
mutexes are just as fast.
- no per-lock kernel allocation happens.
- no resource limits are needed.
- no kernel-space recovery call (FUTEX_RECOVER) is needed.
- the implementation and the locking is "obvious", and there are no
interactions with the VM.
Performance
-----------
I have benchmarked the time needed for the kernel to process a list of 1
million (!) held locks, using the new method [on a 2GHz CPU]:
- with FUTEX_WAIT set [contended mutex]: 130 msecs
- without FUTEX_WAIT set [uncontended mutex]: 30 msecs
I have also measured an approach where glibc does the lock notification
[which it currently does for !pshared robust mutexes], and that took 256
msecs - clearly slower, due to the 1 million FUTEX_WAKE syscalls
userspace had to do.
(1 million held locks are unheard of - we expect at most a handful of
locks to be held at a time. Nevertheless it's nice to know that this
approach scales nicely.)
Implementation details
----------------------
The patch adds two new syscalls: one to register the userspace list, and
one to query the registered list pointer:
asmlinkage long
sys_set_robust_list(struct robust_list_head __user *head,
size_t len);
asmlinkage long
sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
size_t __user *len_ptr);
List registration is very fast: the pointer is simply stored in
current->robust_list. [Note that in the future, if robust futexes become
widespread, we could extend sys_clone() to register a robust-list head
for new threads, without the need of another syscall.]
So there is virtually zero overhead for tasks not using robust futexes,
and even for robust futex users, there is only one extra syscall per
thread lifetime, and the cleanup operation, if it happens, is fast and
straightforward. The kernel doesnt have any internal distinction between
robust and normal futexes.
If a futex is found to be held at exit time, the kernel sets the
following bit of the futex word:
#define FUTEX_OWNER_DIED 0x40000000
and wakes up the next futex waiter (if any). User-space does the rest of
the cleanup.
Otherwise, robust futexes are acquired by glibc by putting the TID into
the futex field atomically. Waiters set the FUTEX_WAITERS bit:
#define FUTEX_WAITERS 0x80000000
and the remaining bits are for the TID.
Testing, architecture support
-----------------------------
i've tested the new syscalls on x86 and x86_64, and have made sure the
parsing of the userspace list is robust [ ;-) ] even if the list is
deliberately corrupted.
i386 and x86_64 syscalls are wired up at the moment, and Ulrich has
tested the new glibc code (on x86_64 and i386), and it works for his
robust-mutex testcases.
All other architectures should build just fine too - but they wont have
the new syscalls yet.
Architectures need to implement the new futex_atomic_cmpxchg_inuser()
inline function before writing up the syscalls (that function returns
-ENOSYS right now).