to tune pthread mutex performance:
1. LIBPTHREAD_SPINLOOPS
If a pthread mutex is being locked by another thread, this environment
variable sets total number of spin loops before the current thread
sleeps in kernel, this saves a syscall overhead if the mutex will be
unlocked very soon (well written application code).
2. LIBPTHREAD_YIELDLOOPS
If a pthread mutex is being locked by other threads, this environment
variable sets total number of sched_yield() loops before the currrent
thread sleeps in kernel. if a pthread mutex is locked, the current thread
gives up cpu, but will not sleep in kernel, this means, current thread
does not set contention bit in mutex, but let lock owner to run again
if the owner is on kernel's run queue, and when lock owner unlocks the
mutex, it does not need to enter kernel and do lots of work to resume
mutex waiters, in some cases, this saves lots of syscall overheads for
mutex owner.
In my practice, sometimes LIBPTHREAD_YIELDLOOPS can massively improve performance
than LIBPTHREAD_SPINLOOPS, this depends on application. These two environments
are global to all pthread mutex, there is no interface to set them for each
pthread mutex, the default values are zero, this means spinning is turned off
by default.
is also implemented in glibc and is used by a number of existing
applications (mysql, firefox, etc).
This mutex type is a default mutex with the additional property that
it spins briefly when attempting to acquire a contested lock, doing
trylock operations in userland before entering the kernel to block if
eventually unsuccessful.
The expectation is that applications requesting this mutex type know
that the mutex is likely to be only held for very brief periods, so it
is faster to spin in userland and probably succeed in acquiring the
mutex, than to enter the kernel and sleep, only to be woken up almost
immediately. This can help significantly in certain cases when
pthread mutexes are heavily contended and held for brief durations
(such as mysql).
Spin up to 200 times before entering the kernel, which represents only
a few us on modern CPUs. No performance degradation was observed with
this value and it is sufficient to avoid a large performance drop in
mysql performance in the heavily contended pthread mutex case.
The libkse implementation is a NOP.
Reviewed by: jeff
MFC after: 3 days
into pthread structure to keep track of locked PTHREAD_PRIO_PROTECT mutex,
no real mutex code is changed, the mutex locking and unlocking code should
has same performance as before.
to make it work, turnstile like mechanism to support priority
propagating and other realtime scheduling options in kernel
should be available to userland mutex, for the moment, I just
want to make libthr be simple and efficient thread library.
Discussed with: deischen, julian
1. fast simple type mutex.
2. __thread tls works.
3. asynchronous cancellation works ( using signal ).
4. thread synchronization is fully based on umtx, mainly, condition
variable and other synchronization objects were rewritten by using
umtx directly. those objects can be shared between processes via
shared memory, it has to change ABI which does not happen yet.
5. default stack size is increased to 1M on 32 bits platform, 2M for
64 bits platform.
As the result, some mysql super-smack benchmarks show performance is
improved massivly.
Okayed by: jeff, mtm, rwatson, scottl
specified mutex is invalid. In spec parlance 'MAY FAIL' means it's
up to the implementor. So, remove the check for NULL pointers for two
reasons:
1. A mutex may be invalid without necessarily being NULL.
2. If the pointer to the mutex is NULL core-dumping in the
vicinity of the problem is much much much better than failing
in some other part of the code (especially when the application
doesn't check the return value of the function that you oh so
helpfully set to EINVAL).
o In the rwlock code: move a duplicated check inside an if..else to after
the if...else clause.
o When initializing a static rwlock move the initialization check
inside the lock.
o In thr_setschedparam.c: When breaking out of the trylock...retry if busy
loop make sure to reset the mtx pointer to null if the mutex is nolonger
in a queue.
pointer to the corresponding struct thread to the thread ID (lwpid_t)
assigned to that thread. The primary reason for this change is that
libthr now internally uses the same ID as the debugger and the kernel
when referencing to a kernel thread. This allows us to implement the
support for debugging without additional translations and/or mappings.
To preserve the ABI, the 1:1 threading syscalls, including the umtx
locking API have not been changed to work on a lwpid_t. Instead the
1:1 threading syscalls operate on long and the umtx locking API has
not been changed except for the contested bit. Previously this was
the least significant bit. Now it's the most significant bit. Since
the contested bit should not be tested by userland, this change is
not expected to be visible. Just to be sure, UMTX_CONTESTED has been
removed from <sys/umtx.h>.
Reviewed by: mtm@
ABI preservation tested on: i386, ia64
followed are: Only 3 functions (pthread_cancel, pthread_setcancelstate,
pthread_setcanceltype) are required to be async-signal-safe by POSIX. None of
the rest of the pthread api is required to be async-signal-safe. This means
that only the three mentioned functions are safe to use from inside
signal handlers.
However, there are certain system/libc calls that are
cancellation points that a caller may call from within a signal handler,
and since they are cancellation points calls have to be made into libthr
to test for cancellation and exit the thread if necessary. So, the
cancellation test and thread exit code paths must be async-signal-safe
as well. A summary of the changes follows:
o Almost all of the code paths that masked signals, as well as locking the
pthread structure now lock only the pthread structure.
o Signals are masked (and left that way) as soon as a thread enters
pthread_exit().
o The active and dead threads locks now explicitly require that signals
are masked.
o Access to the isdead field of the pthread structure is protected by both
the active and dead list locks for writing. Either one is sufficient for
reading.
o The thread state and type fields have been combined into one three-state
switch to make it easier to read without requiring a lock. It doesn't need
a lock for writing (and therefore for reading either) because only the
current thread can write to it and it is an integer value.
o The thread state field of the pthread structure has been eliminated. It
was an unnecessary field that mostly duplicated the flags field, but
required additional locking that would make a lot more code paths require
signal masking. Any truly unique values (such as PS_DEAD) have been
reborn as separate members of the pthread structure.
o Since the mutex and condvar pthread functions are not async-signal-safe
there is no need to muck about with the wait queues when handling
a signal ...
o ... which also removes the need for wrapping signal handlers and sigaction(2).
o The condvar and mutex async-cancellation code had to be revised as a result
of some of these changes, which resulted in semi-unrelated changes which
would have been difficult to work on as a separate commit, so they are
included as well.
The only part of the changes I am worried about is related to locking for
the pthread joining fields. But, I will take a closer look at them once this
mega-patch is committed.
we still have to DTRT when an asynchronously cancellable thread is
cancelled while waiting for a mutex.
o While dequeueing a waiting mutex don't skip a thread if it has
a cancel pending. Only skip it if it is also async cancellable.
o Fix mutex priority protocols. Keep separate counts of priority
inheritance and protection mutexes to make things easier.
This will not have much affect since this is only the
userland side, and the rest involves kernel scheduling.
o Simplify the logic by removing a lot of unnecesary nesting
o Reduce the amount of local variables
o Zero-out the allocated structure and get rid of
all the unnecessary setting to 0 and NULL;
Refactor _pthread_mutex_destroy
o Simplify the logic by removing a lot of unnecesary nesting
o No need to check pointer that the mutex attributes points
to. Checking passed in pointer is enough.
waiting on a locked mutex. This involves passing a struct timespec
from the pthread mutex locking interfaces all the way down to the
function that suspends the thread until the mutex is released.
The timeout is assumed to be an absolute time (i.e. not relative to
the current time).
Also, in _thread_suspend() make the passed in timespec const.
Create a private, single underscore, version of pthread_mutex_unlock for libc.
pthread_mutex_lock already has one. These versions are different from the
ones that applications will link against because they block all signals
from the time a call to lock the mutex is made until it is successfully
unlocked.
a thread receives a spurious wakeup from sigtimedwait(), so make sure
that the call to the queueing code is called only once before entering
the loop (not in the loop). This should fix some fatal errors people
are seeing with messages stating the thread is already on the mutex queue.
These errors may still be triggered from signal handlers; however, since
that part of the code is not locked down yet.
