[PATCH] lightweight robust futexes: docs

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.

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).