Lock elision using TSX is a technique to optimize lock scaling
It allows to run locks in parallel using hardware support for
a transactional execution mode in 4th generation Intel Core CPUs.
See http://www.intel.com/software/tsx for more Information.
This patch implements a simple adaptive lock elision algorithm based
on RTM. It enables elision for the pthread mutexes and rwlocks.
The algorithm keeps track whether a mutex successfully elides or not,
and stops eliding for some time when it is not.
When the CPU supports RTM the elision path is automatically tried,
otherwise any elision is disabled.
The adaptation algorithm and its tuning is currently preliminary.
The code adds some checks to the lock fast paths. Micro-benchmarks
show little to no difference without RTM.
This patch implements the low level "lll_" code for lock elision.
Followon patches hook this into the pthread implementation
Changes with the RTM mutexes:
-----------------------------
Lock elision in pthreads is generally compatible with existing programs.
There are some obscure exceptions, which are expected to be uncommon.
See the manual for more details.
- A broken program that unlocks a free lock will crash.
There are ways around this with some tradeoffs (more code in hot paths)
I'm still undecided on what approach to take here; have to wait for testing reports.
- pthread_mutex_destroy of a lock mutex will not return EBUSY but 0.
- There's also a similar situation with trylock outside the mutex,
"knowing" that the mutex must be held due to some other condition.
In this case an assert failure cannot be recovered. This situation is
usually an existing bug in the program.
- Same applies to the rwlocks. Some of the return values changes
(for example there is no EDEADLK for an elided lock, unless it aborts.
However when elided it will also never deadlock of course)
- Timing changes, so broken programs that make assumptions about specific timing
may expose already existing latent problems. Note that these broken programs will
break in other situations too (loaded system, new faster hardware, compiler
optimizations etc.)
- Programs with non recursive mutexes that take them recursively in a thread and
which would always deadlock without elision may not always see a deadlock.
The deadlock will only happen on an early or delayed abort (which typically
happens at some point)
This only happens for mutexes not explicitely set to PTHREAD_MUTEX_NORMAL
or PTHREAD_MUTEX_ADAPTIVE_NP. PTHREAD_MUTEX_NORMAL mutexes do not elide.
The elision default can be set at configure time.
This patch implements the basic infrastructure for elision.
Resolves BZ #15618.
pthread_attr_getaffinity_np may write beyond bounds of the input
cpuset buffer if the size of the input buffer is smaller than the
buffer present in the input pthread attributes. Fix is to copy to the
extent of the minimum of the source and the destination.
It is very very possible that the futex syscall returns an
error and that the caller of lll_futex_wake may want to
look at that error and propagate the failure.
This patch allows a caller to see the syscall error.
There are no users of the syscall error at present, but
future cleanups are now be able to check for the error.
--
nplt/
2013-06-10 Carlos O'Donell <carlos@redhat.com>
* sysdeps/unix/sysv/linux/i386/lowlevellock.h
(lll_futex_wake): Return syscall error.
* sysdeps/unix/sysv/linux/x86_64/lowlevellock.h
(lll_futex_wake): Return syscall error.
The sem_post.c file uses atomic functions without including
atomic.h. Add `#include <atomic.h>' to the file to prevent
any compile time warnings when other headers change and
atomic.h isn't implicitly included.
---
nptl/
2013-04-07 Carlos O'Donell <carlos@redhat.com>
* sysdeps/unix/sysv/linux/sem_post.c: Include atomic.h.
Fixes BZ #15337.
Static builds fail with the following warning:
/home/tools/glibc/glibc/nptl/../nptl/sysdeps/unix/sysv/linux/x86_64/cancellation.S:80:
undefined reference to `__GI___pthread_unwind'
When the source is configured with --disable-hidden-plt. This is
because the preprocessor conditional in cancellation.S only checks if
the build is for SHARED, whereas hidden_def is defined appropriately
only for a SHARED build that will have symbol versioning *and* hidden
defs are enabled. The last case is false here.
* sysdeps/unix/sysv/linux/i386/i486/pthread_cond_timedwait.S
(__pthread_cond_timedwait): If possible use FUTEX_WAIT_BITSET to
directly use absolute timeout.
Since the FUTEX_WAIT operation takes a relative timeout, the
pthread_cond_timedwait and other timed function implementations have
to get a relative timeout from the absolute timeout parameter it gets
before it makes the futex syscall. This value is then converted back
into an absolute timeout within the kernel. This is a waste and has
hence been improved upon by a FUTEX_WAIT_BITSET operation (OR'd with
FUTEX_CLOCK_REALTIME to make the kernel use the realtime clock instead
of the default monotonic clock). This was implemented only in the x86
and sh assembly code and not in the C code. This patch implements
support for FUTEX_WAIT_BITSET whenever available (since linux-2.6.29)
for s390 and powerpc.
nptl/
* sysdeps/unix/sysv/linux/sparc/lowlevellock.h (BUSY_WAIT_NOP):
Define when we have v9 instructions available.
* sysdeps/unix/sysv/linux/sparc/sparc64/cpu_relax.S: New file.
* sysdeps/unix/sysv/linux/sparc/sparc32/sparcv9/cpu_relax.S: New
file.
* sysdeps/unix/sysv/linux/sparc/sparc32/sparcv9/Makefile: New
file.
* sysdeps/unix/sysv/linux/sparc/sparc64/Makefile: Add cpu_relax
to libpthread-routines.
[BZ #14652]
When a thread waiting in pthread_cond_wait with a PI mutex is
cancelled after it has returned successfully from the futex syscall
but just before async cancellation is disabled, it enters its
cancellation handler with the mutex held and simply calling a
mutex_lock again will result in a deadlock. Hence, it is necessary to
see if the thread owns the lock and try to lock it only if it doesn't.
[BZ #14417]
A futex call with FUTEX_WAIT_REQUEUE_PI returns with the mutex locked
on success. If such a successful thread is pipped to the cond_lock by
another spuriously woken waiter, it could be sent back to wait on the
futex with the mutex lock held, thus causing a deadlock. So it is
necessary that the thread relinquishes the mutex before going back to
sleep.
[BZ #14477]
Add an additional entry in the exception table to jump to
__condvar_w_cleanup2 instead of __condvar_w_cleanup for PI mutexes
when %ebx contains the address of the futex instead of the condition
variable.
Ref gcc.gnu.org/bugzilla/show_bug.cgi?id=52839#c10
Release barriers are needed to ensure that any memory written by
init_routine is seen by other threads before *once_control changes.
In the case of clear_once_control we need to flush any partially
written state.
There is nothing in the POSIX specification to disallow a
single-threaded program from cancelling itself, so we forcibly enable
multiple_threads to allow the next available cancellation point in the
thread to run.
Also added additional tests to cover various cancellation scenarios.
nptl/
[BZ #13844]
* sysdeps/unix/sysv/linux/libc-lowlevellock.c: Include using <..>
instead of "...".
* sysdeps/unix/sysv/linux/sparc/sparc32/libc-lowlevellock.c:
Delete, not needed.
/
* sysdeps/unix/sysv/linux/sparc/sparc64/get_clockfreq.c
(set_obp_int): New function.
(get_obp_int): New function.
(__get_clockfreq_via_dev_openprom): Likewise.
* sysdeps/unix/sysv/linux/sparc/sysdep.h (INTERNAL_SYSCALL_ERROR_P): Avoid
unused variable warnings on 'val' and use builtin_expect.
(INLINE_SYSCALL): Don't wrap INTERNAL_SYSCALL_ERROR_P with builtin_expect.
(INLINE_CLONE_SYSCALL): Likewise.
nptl/
* sysdeps/unix/sysv/linux/sparc/sem_post.c (__new_sem_post): Use
atomic_increment and remove unused local variable.
(__old_sem_post): Likewise.
Andi Kleen
Siddhesh Poyarekar
Carlos O'Donell
Joseph Myers
Andreas Schwab
Roland McGrath
David S. Miller
Alan Modra
Andreas Schwab
Thomas Schwinge
H.J. Lu
Chung-Lin Tang
Rayson Ho
Andreas Jaeger
Ulrich Drepper
Richard Henderson
Kaz Kojima
Marek Polacek
Paul Eggert
Adhemerval Zanella
Andreas Schwab
H.J. Lu
Jim Meyering
Dinakar Guniguntala