critical_enter() and critical_exit() are now solely a mechanism for
deferring kernel preemptions. They no longer have any affect on
interrupts. This means that standalone critical sections are now very
cheap as they are simply unlocked integer increments and decrements for the
common case.
Spin mutexes now use a separate KPI implemented in MD code: spinlock_enter()
and spinlock_exit(). This KPI is responsible for providing whatever MD
guarantees are needed to ensure that a thread holding a spin lock won't
be preempted by any other code that will try to lock the same lock. For
now all archs continue to block interrupts in a "spinlock section" as they
did formerly in all critical sections. Note that I've also taken this
opportunity to push a few things into MD code rather than MI. For example,
critical_fork_exit() no longer exists. Instead, MD code ensures that new
threads have the correct state when they are created. Also, we no longer
try to fixup the idlethreads for APs in MI code. Instead, each arch sets
the initial curthread and adjusts the state of the idle thread it borrows
in order to perform the initial context switch.
This change is largely a big NOP, but the cleaner separation it provides
will allow for more efficient alternative locking schemes in other parts
of the kernel (bare critical sections rather than per-CPU spin mutexes
for per-CPU data for example).
Reviewed by: grehan, cognet, arch@, others
Tested on: i386, alpha, sparc64, powerpc, arm, possibly more
turn it back on. Specifically, the actual changes are now less intrusive
in that the _get_spin_lock() and _rel_spin_lock() macros now have their
contents changed for UP vs SMP kernels which centralizes the changes.
Also, UP kernels do not use _mtx_lock_spin() and no longer include it. The
UP versions of the spin lock functions do not use any atomic operations,
but simple compares and stores which allow mtx_owned() to still work for
spin locks while removing the overhead of atomic operations.
Tested on: i386, alpha
- Add a new _lock() call to each API that locks the associated chain lock
for a lock_object pointer or wait channel. The _lookup() functions now
require that the chain lock be locked via _lock() when they are called.
- Change sleepq_add(), turnstile_wait() and turnstile_claim() to lookup
the associated queue structure internally via _lookup() rather than
accepting a pointer from the caller. For turnstiles, this means that
the actual lookup of the turnstile in the hash table is only done when
the thread actually blocks rather than being done on each loop iteration
in _mtx_lock_sleep(). For sleep queues, this means that sleepq_lookup()
is no longer used outside of the sleep queue code except to implement an
assertion in cv_destroy().
- Change sleepq_broadcast() and sleepq_signal() to require that the chain
lock is already required. For condition variables, this lets the
cv_broadcast() and cv_signal() functions lock the sleep queue chain lock
while testing the waiters count. This means that the waiters count
internal to condition variables is no longer protected by the interlock
mutex and cv_broadcast() and cv_signal() now no longer require that the
interlock be held when they are called. This lets consumers of condition
variables drop the lock before waking other threads which can result in
fewer context switches.
MFC after: 1 month
FULL_PREEMPTION is defined. Add a runtime warning to ULE if PREEMPTION is
enabled (code inspired by the PREEMPTION warning in kern_switch.c). This
is a possible MT5 candidate.
macros and pass the value to the associated _mtx_*() functions to avoid
more curthread dereferences in the function implementations. This provided
a very modest perf improvement in some benchmarks.
Suggested by: rwatson
Tested by: scottl
being defined, define and use a new MD macro, cpu_spinwait(). It only
expands to something on i386 and amd64, so the compiled code should be
identical.
Name of the macro found by: jhb
Reviewed by: jhb
an adaptive fashion when adaptive mutexes are enabled. The theory
behind non-adaptive Giant is that Giant will be held for long periods
of time, and therefore spinning waiting on it is wasteful. However,
in MySQL benchmarks which are relatively Giant-free, running Giant
adaptive makes an observable difference on SMP (5% transaction rate
improvement). As such, make adaptive behavior on Giant an option so
it can be more widely benchmarked.
NO_ADAPTIVE_MUTEXES. This option has been enabled by default on amd64 for
quite some time, and has been extensively tested on i386 and sparc64. It
shows measurable performance gains in many circumstances, and few negative
effects. It would be nice in t he future if adaptive mutexes actually went
to sleep after a certain amount of spinning, but that will require quite a
bit more testing.
o Make debugging code conditional upon KDB instead of DDB.
o Call kdb_enter() instead of Debugger().
o Call kdb_backtrace() instead of db_print_backtrace() or backtrace().
kern_mutex.c:
o Replace checks for db_active with checks for kdb_active and make
them unconditional.
kern_shutdown.c:
o s/DDB_UNATTENDED/KDB_UNATTENDED/g
o s/DDB_TRACE/KDB_TRACE/g
o Save the TID of the thread doing the kernel dump so the debugger
knows which thread to select as the current when debugging the
kernel core file.
o Clear kdb_active instead of db_active and do so unconditionally.
o Remove backtrace() implementation.
kern_synch.c:
o Call kdb_reenter() instead of db_error().
than as one-off hacks in various other parts of the kernel:
- Add a function maybe_preempt() that is called from sched_add() to
determine if a thread about to be added to a run queue should be
preempted to directly. If it is not safe to preempt or if the new
thread does not have a high enough priority, then the function returns
false and sched_add() adds the thread to the run queue. If the thread
should be preempted to but the current thread is in a nested critical
section, then the flag TDF_OWEPREEMPT is set and the thread is added
to the run queue. Otherwise, mi_switch() is called immediately and the
thread is never added to the run queue since it is switch to directly.
When exiting an outermost critical section, if TDF_OWEPREEMPT is set,
then clear it and call mi_switch() to perform the deferred preemption.
- Remove explicit preemption from ithread_schedule() as calling
setrunqueue() now does all the correct work. This also removes the
do_switch argument from ithread_schedule().
- Do not use the manual preemption code in mtx_unlock if the architecture
supports native preemption.
- Don't call mi_switch() in a loop during shutdown to give ithreads a
chance to run if the architecture supports native preemption since
the ithreads will just preempt DELAY().
- Don't call mi_switch() from the page zeroing idle thread for
architectures that support native preemption as it is unnecessary.
- Native preemption is enabled on the same archs that supported ithread
preemption, namely alpha, i386, and amd64.
This change should largely be a NOP for the default case as committed
except that we will do fewer context switches in a few cases and will
avoid the run queues completely when preempting.
Approved by: scottl (with his re@ hat)
switch to. If a non-NULL thread pointer is passed in, then the CPU will
switch to that thread directly rather than calling choosethread() to pick
a thread to choose to.
- Make sched_switch() aware of idle threads and know to do
TD_SET_CAN_RUN() instead of sticking them on the run queue rather than
requiring all callers of mi_switch() to know to do this if they can be
called from an idlethread.
- Move constants for arguments to mi_switch() and thread_single() out of
the middle of the function prototypes and up above into their own
section.
to awaken all waiters when a contested mutex is released instead of just
the highest priority waiter. If the various threads are awakened in
sequence then each thread may acquire and release the lock in question
without contention resulting in fewer expensive unlock and lock
operations. This old behavior of waking just the highest priority is
still used if this option is specified. Making the algorithm conditional
on a kernel option will allows us to benchmark both cases later and
determine which one should be used by default.
Requested by: tanimura-san
profiling buffers and hash table. This makes it a lot easier to
do multiple profiling runs without rebooting or performing
gratuitous arithmetic. Sysctl is named debug.mutex.prof.reset.
Reviewed by: jake
- witness_lock() is split into two pieces: witness_checkorder() and
witness_lock(). Witness_checkorder() determines if acquiring a specified
lock at the time it is called would result in a lock order. It
optionally adds a new lock order relationship as well. witness_lock()
updates witness's data structures to assume that a lock has been acquired
by stick a new lock instance in the appropriate lock instance list.
- The mutex and sx lock functions now call checkorder() prior to trying to
acquire a lock and continue to call witness_lock() after the acquire is
completed. This will let witness catch a deadlock before it happens
rather than trying to do so after the threads have deadlocked (i.e. never
actually report it).
- A new function witness_defineorder() has been added that adds a lock
order between two locks at runtime without having to acquire the locks.
If the lock order cannot be added it will return an error. This function
is available to programmers via the WITNESS_DEFINEORDER() macro which
accepts either two mutexes or two sx locks as its arguments.
- A few simple wrapper macros were added to allow developers to call
witness_checkorder() anywhere as a way of enforcing locking assertions
in code that might acquire a certain lock in some situations. The
macros are: witness_check_{mutex,shared_sx,exclusive_sx} and take an
appropriate lock as the sole argument.
- The code to remove a lock instance from a lock list in witness_unlock()
was unnested by using a goto to vastly improve the readability of this
function.
SW_INVOL. Assert that one of these is set in mi_switch() and propery
adjust the rusage statistics. This is to simplify the large number of
users of this interface which were previously all required to adjust the
proper counter prior to calling mi_switch(). This also facilitates more
switch and locking optimizations.
- Change all callers of mi_switch() to pass the appropriate paramter and
remove direct references to the process statistics.
mutex profiling code. As with existing mutex profiling, measurement
is done with respect to mtx_lock() instances in the code, as opposed
to specific mutexes. In particular, measure two things:
(1) Lock contention. How often did this mtx_lock() call get made and
have to sleep (or almost sleep) waiting for the lock. This helps
identify the "victims" of contention.
(2) Hold contention. How often, while the lock was held by a thread
as a result of this mtx_lock(), did another thread try to acquire
the same mutex. This helps identify the causes of contention.
I'm currently exploring adding measurement of "time waited for the
lock", but the current implementation has proven useful to me so far
so I figured I'd commit it so others could try it out. Note that this
increases the size of mutexes when MUTEX_PROFILING is enabled, so you
might find you need to further bump UMA_BOOT_PAGES. Fixes welcome.
The once over: des, others
on a non-recursive mutex will fail but will not trigger any assertions.
- Add an assertion to mtx_lock() that one never recurses on a non-recursive
mutex. This is mostly useful for the non-WITNESS case.
Requested by: deischen, julian, others (1)
turnstiles to implement blocking isntead of implementing a thread queue
directly. These turnstiles are somewhat similar to those used in Solaris 7
as described in Solaris Internals but are also different.
Turnstiles do not come out of a fixed-sized pool. Rather, each thread is
assigned a turnstile when it is created that it frees when it is destroyed.
When a thread blocks on a lock, it donates its turnstile to that lock to
serve as queue of blocked threads. The queue associated with a given lock
is found by a lookup in a simple hash table. The turnstile itself is
protected by a lock associated with its entry in the hash table. This
means that sched_lock is no longer needed to contest on a mutex. Instead,
sched_lock is only used when manipulating run queues or thread priorities.
Turnstiles also implement priority propagation inherently.
Currently turnstiles only support mutexes. Eventually, however, turnstiles
may grow two queue's to support a non-sleepable reader/writer lock
implementation. For more details, see the comments in sys/turnstile.h and
kern/subr_turnstile.c.
The two primary advantages from the turnstile code include: 1) the size
of struct mutex shrinks by four pointers as it no longer stores the
thread queue linkages directly, and 2) less contention on sched_lock in
SMP systems including the ability for multiple CPUs to contend on different
locks simultaneously (not that this last detail is necessarily that much of
a big win). Note that 1) means that this commit is a kernel ABI breaker,
so don't mix old modules with a new kernel and vice versa.
Tested on: i386 SMP, sparc64 SMP, alpha SMP
thread's pid to make debugging easier for people who don't want to have to
use the intended tool for these panics (witness).
Indirectly prodded by: kris
locks held by each thread.
- Fix a bug in the original BSD/OS code where a contested lock was not
properly handed off from the old thread to the new thread when a
contested lock with more than one blocked thread was transferred from
one thread to another.
- Don't use an atomic operation to write the MTX_CONTESTED value to
mtx_lock in the aforementioned special case. The memory barriers and
exclusion provided by sched_lock are sufficient.
Spotted by: alc (2)
already own. The mtx_trylock() will fail however. Enhance the comment
at the top of the try lock function to explain this.
Requested by: jlemon and his evil netisr locking
- Declare some local variables at the top of the function instead of in a
nested block.
- Use mtx_owned() instead of masking off bits from mtx_lock manually.
- Read the value of mtx_lock into 'v' as a separate line rather than inside
an if statement for clarity. This code is hairy enough as it is.
owned. Previously the KASSERT would only trigger if we successfully
acquired a lock that we already held. However, _obtain_lock() fails to
acquire locks that we already hold, so the KASSERT was never checked in
the case it was supposed to fail.
o Always check for null when dereferencing the filename component.
o Implement a try-and-backoff method for allocating memory to
dump stats to avoid a spin-lock -> sleep-lock mutex lock order
panic with WITNESS.
Approved by: des, markm (mentor)
Not objected: jhb
does not require Giant.
This means that we may miss panics on a class of mutex programming bugs,
but only if running with a Chernobyl setting of debug-flags.
Spotted by: Pete Carah <pete@ns.altadena.net>
- Begin moving scheduler specific functionality into sched_4bsd.c
- Replace direct manipulation of scheduler data with hooks provided by the
new api.
- Remove KSE specific state modifications and single runq assumptions from
kern_switch.c
Reviewed by: -arch
name instead. (e.g., SLOCK instead of SMTX, TD_ON_LOCK() instead of
TD_ON_MUTEX()) Eventually a turnstile abstraction will be added that
will be shared with mutexes and other types of locks. SLOCK/TDI_LOCK will
be used internally by the turnstile code and will not be specific to
mutexes. Making the change now ensures that turnstiles can be dropped
in at a later date without affecting the ABI of userland applications.
from stopping another thread from completing a syscall, and this allows it to
release its resources etc. Probably more related commits to follow (at least
one I know of)
Initial concept by: julian, dillon
Submitted by: davidxu
sleep mutexes and vice versa. WITNESS normally should catch this but
not everyone uses WITNESS so this is a fallback to catch nasty but easy
to do bugs.
log the start and end of periods during which mtx_lock() is waiting
to acquire a sleep mutex. The log message includes the file and
line of both the waiter and the holder.
Reviewed by: jhb, jake
