Historical behavior of letting other CPUs merily go on is a default for
time being. The new behavior can be switched on via
kern.stop_scheduler_on_panic tunable and sysctl.
Stopping of the CPUs has (at least) the following benefits:
- more of the system state at panic time is preserved intact
- threads and interrupts do not interfere with dumping of the system
state
Only one thread runs uninterrupted after panic if stop_scheduler_on_panic
is set. That thread might call code that is also used in normal context
and that code might use locks to prevent concurrent execution of certain
parts. Those locks might be held by the stopped threads and would never
be released. To work around this issue, it was decided that instead of
explicit checks for panic context, we would rather put those checks
inside the locking primitives.
This change has substantial portions written and re-written by attilio
and kib at various times. Other changes are heavily based on the ideas
and patches submitted by jhb and mdf. bde has provided many insights
into the details and history of the current code.
The new behavior may cause problems for systems that use a USB keyboard
for interfacing with system console. This is because of some unusual
locking patterns in the ukbd code which have to be used because on one
hand ukbd is below syscons, but on the other hand it has to interface
with other usb code that uses regular mutexes/Giant for its concurrency
protection. Dumping to USB-connected disks may also be affected.
PR: amd64/139614 (at least)
In cooperation with: attilio, jhb, kib, mdf
Discussed with: arch@, bde
Tested by: Eugene Grosbein <eugen@grosbein.net>,
gnn,
Steven Hartland <killing@multiplay.co.uk>,
glebius,
Andrew Boyer <aboyer@averesystems.com>
(various versions of the patch)
MFC after: 3 months (or never)
This enables locking consumers to pass their own structures around as const and
be able to assert locks embedded into those structures.
Reviewed by: ed, kib, jhb
The SYSCTL_NODE macro defines a list that stores all child-elements of
that node. If there's no SYSCTL_DECL macro anywhere else, there's no
reason why it shouldn't be static.
sxlock, via the sx_{s, x}lock_sig() interface, or plain lockmgr), will
leave the waiters flag on forcing the owner to do a wakeup even when if
the waiter queue is empty.
That operation may lead to a deadlock in the case of doing a fake wakeup
on the "preferred" (based on the wakeup algorithm) queue while the other
queue has real waiters on it, because nobody is going to wakeup the 2nd
queue waiters and they will sleep indefinitively.
A similar bug, is present, for lockmgr in the case the waiters are
sleeping with LK_SLEEPFAIL on. In this case, even if the waiters queue
is not empty, the waiters won't progress after being awake but they will
just fail, still not taking care of the 2nd queue waiters (as instead the
lock owned doing the wakeup would expect).
In order to fix this bug in a cheap way (without adding too much locking
and complicating too much the semantic) add a sleepqueue interface which
does report the actual number of waiters on a specified queue of a
waitchannel (sleepq_sleepcnt()) and use it in order to determine if the
exclusive waiters (or shared waiters) are actually present on the lockmgr
(or sx) before to give them precedence in the wakeup algorithm.
This fix alone, however doesn't solve the LK_SLEEPFAIL bug. In order to
cope with it, add the tracking of how many exclusive LK_SLEEPFAIL waiters
a lockmgr has and if all the waiters on the exclusive waiters queue are
LK_SLEEPFAIL just wake both queues.
The sleepq_sleepcnt() introduction and ABI breakage require
__FreeBSD_version bumping.
Reported by: avg, kib, pho
Reviewed by: kib
Tested by: pho
in order to avoid, on architectures which doesn't have strong ordered
writes, CPU instructions reordering.
Diagnosed by: fabio
Reviewed by: jhb
Tested by: Giovanni Trematerra
<giovanni dot trematerra at gmail dot com>
In the lockmgr support:
- GIANT_RESTORE() is just called when the sleep finishes, so the current
code can ends up into a giant unlock problem. Fix it by appropriately
call GIANT_RESTORE() when needed. Note that this is not exactly ideal
because for any interation of the adaptive spinning we drop and restore
Giant, but the overhead should be not a factor.
- In the lock held in exclusive mode case, after the adaptive spinning is
brought to completition, we should just retry to acquire the lock
instead to fallthrough. Fix that.
- Fix a style nit
In the sx support:
- Call GIANT_SAVE() before than looping. This saves some overhead because
in the current code GIANT_SAVE() is called several times.
Tested by: Giovanni Trematerra <giovanni dot trematerra at gmail dot com>
a pointer-fetching specific operation check. Consequently, rename the
operation ASSERT_ATOMIC_LOAD_PTR().
* Fix the implementation of ASSERT_ATOMIC_LOAD_PTR() by checking
directly alignment on the word boundry, for all the given specific
architectures. That's a bit too strict for some common case, but it
assures safety.
* Add a comment explaining the scope of the macro
* Add a new stub in the lockmgr specific implementation
Tested by: marcel (initial version), marius
Reviewed by: rwatson, jhb (comment specific review)
Approved by: re (kib)
Check that the given variable is at most uintptr_t in size and that
it is aligned.
Note: ASSERT_ATOMIC_LOAD() uses ALIGN() to check for adequate
alignment -- however, the function of ALIGN() is to guarantee
alignment, and therefore may lead to stronger alignment
enforcement than necessary for types that are smaller than
sizeof(uintptr_t).
Add checks to mtx, rw and sx locks init functions to detect possible
breakage. This was used during debugging of the problem fixed with
r196118 where a pointer was on an un-aligned address in the dpcpu area.
In collaboration with: rwatson
Reviewed by: rwatson
Approved by: re (kib)
Introduce for this operation the reverse NO_ADAPTIVE_SX option.
The flag SX_ADAPTIVESPIN to be passed to sx_init_flags(9) gets suppressed
and the new flag, offering the reversed logic, SX_NOADAPTIVE is added.
Additively implements adaptive spininning for sx held in shared mode.
The spinning limit can be handled through sysctls in order to be tuned
while the code doesn't reach the release, after which time they should
be dropped probabilly.
This change has made been necessary by recent benchmarks where it does
improve concurrency of workloads in presence of high contention
(ie. ZFS).
KPI breakage is documented by __FreeBSD_version bumping, manpage and
UPDATING updates.
Requested by: jeff, kmacy
Reviewed by: jeff
Tested by: pho
adds probes for mutexes, reader/writer and shared/exclusive locks to
gather contention statistics and other locking information for
dtrace scripts, the lockstat(1M) command and other potential
consumers.
Reviewed by: attilio jhb jb
Approved by: gnn (mentor)
of spurious witness warnings since lockmgr grew witness support. Before
this, every time you passed an interlock to a lockmgr lock WITNESS treated
it as a LOR.
Reviewed by: attilio
routine wakes up proc0 so that proc0 can swap the thread back in.
Historically, this has been done by waking up proc0 directly from
setrunnable() itself via a wakeup(). When waking up a sleeping thread
that was swapped out (the usual case when waking proc0 since only sleeping
threads are eligible to be swapped out), this resulted in a bit of
recursion (e.g. wakeup() -> setrunnable() -> wakeup()).
With sleep queues having separate locks in 6.x and later, this caused a
spin lock LOR (sleepq lock -> sched_lock/thread lock -> sleepq lock).
An attempt was made to fix this in 7.0 by making the proc0 wakeup use
the ithread mechanism for doing the wakeup. However, this required
grabbing proc0's thread lock to perform the wakeup. If proc0 was asleep
elsewhere in the kernel (e.g. waiting for disk I/O), then this degenerated
into the same LOR since the thread lock would be some other sleepq lock.
Fix this by deferring the wakeup of the swapper until after the sleepq
lock held by the upper layer has been locked. The setrunnable() routine
now returns a boolean value to indicate whether or not proc0 needs to be
woken up. The end result is that consumers of the sleepq API such as
*sleep/wakeup, condition variables, sx locks, and lockmgr, have to wakeup
proc0 if they get a non-zero return value from sleepq_abort(),
sleepq_broadcast(), or sleepq_signal().
Discussed with: jeff
Glanced at by: sam
Tested by: Jurgen Weber jurgen - ish com au
MFC after: 2 weeks
lock_object, using an unified field called lo_data.
- Replace lo_type usage with the w_name usage and at init time pass the
lock "type" directly to witness_init() from the parent lock init
function. Handle delayed initialization before than
witness_initialize() is called through the witness_pendhelp structure.
- Axe out LO_ENROLLPEND as it is not really needed. The case where the
mutex init delayed wants to be destroyed can't happen because
witness_destroy() checks for witness_cold and panic in case.
- In enroll(), if we cannot allocate a new object from the freelist,
notify that to userspace through a printf().
- Modify the depart function in order to return nothing as in the current
CVS version it always returns true and adjust callers accordingly.
- Fix the witness_addgraph() argument name prototype.
- Remove unuseful code from itismychild().
This commit leads to a shrinked struct lock_object and so smaller locks,
in particular on amd64 where 2 uintptr_t (16 bytes per-primitive) are
gained.
Reviewed by: jhb
sched_sleep(). This removes extra thread_lock() acquisition and
allows the scheduler to decide what to do with the static boost.
- Change the priority arguments to cv_* to match sleepq/msleep/etc.
where 0 means no priority change. Catch -1 in cv_broadcastpri() and
convert it to 0 for now.
- Set a flag when sleeping in a way that is compatible with swapping
since direct priority comparisons are meaningless now.
- Add a sysctl to ule, kern.sched.static_boost, that defaults to on which
controls the boost behavior. Turning it off gives better performance
in some workloads but needs more investigation.
- While we're modifying sleepq, change signal and broadcast to both
return with the lock held as the lock was held on enter.
Reviewed by: jhb, peter
the ABI when enabled. There is no longer an embedded lock_profile_object
in each lock. Instead a list of lock_profile_objects is kept per-thread
for each lock it may own. The cnt_hold statistic is now always 0 to
facilitate this.
- Support shared locking by tracking individual lock instances and
statistics in the per-thread per-instance lock_profile_object.
- Make the lock profiling hash table a per-cpu singly linked list with a
per-cpu static lock_prof allocator. This removes the need for an array
of spinlocks and reduces cache contention between cores.
- Use a seperate hash for spinlocks and other locks so that only a
critical_enter() is required and not a spinlock_enter() to modify the
per-cpu tables.
- Count time spent spinning in the lock statistics.
- Remove the LOCK_PROFILE_SHARED option as it is always supported now.
- Specifically drop and release the scheduler locks in both schedulers
since we track owners now.
In collaboration with: Kip Macy
Sponsored by: Nokia
an unified way for all the lock primitives to express lock assertions.
Currenty, lockmgrs and rmlocks don't have assertions, so just panic in
that case.
This will be a base for more callout improvements.
Ok'ed by: jhb, jeff
opposed to what process. Since threads by default have teh name of the
process unless over-written with more useful information, just print the
thread name instead.
Before that fix, it was possible for the function to fail if number
of sharers changes between 'x = sx->sx_lock' step and atomic_cmpset_acq_ptr()
call.
This fixes ZFS problem when ZFS returns strange EIO errors under load.
In ZFS there is a code that depends on the fact that sx_try_slock() can
only fail if there is an exclusive owner.
Discussed with: attilio
Reviewed by: jhb
Approved by: re (kensmith)
- Adjust lock_profiling stubs semantic in the hard functions in order to be
more accurate and trustable
- Disable shared paths for lock_profiling. Actually, lock_profiling has a
subtle race which makes results caming from shared paths not completely
trustable. A macro stub (LOCK_PROFILING_SHARED) can be actually used for
re-enabling this paths, but is currently intended for developing use only.
- Use homogeneous names for automatic variables in hard functions regarding
lock_profiling
- Style fixes
- Add a CTASSERT for some flags building
Discussed with: kmacy, kris
Approved by: jeff (mentor)
Approved by: re
These functions are intended to do the same actions of sx_xlock() and
sx_slock() but with the difference to perform an interruptible sleep, so
that sleep can be interrupted by external events.
In order to support these new featueres, some code renstruction is needed,
but external API won't be affected at all.
Note: use "void" cast for "int" returning functions in order to avoid tools
like Coverity prevents to whine.
Requested by: rwatson
Tested by: rwatson
Reviewed by: jhb
Approved by: jeff (mentor)
"0" cannot be a correct value since when the function is entered at least
one shared holder must be present and since we want the last one "1" is
the correct value.
Note that lock_profiling for sx locks is far from being perfect.
Expect further fixes for that.
Approved by: jeff (mentor)
obtaining and releasing shared and exclusive locks. The algorithms for
manipulating the lock cookie are very similar to that rwlocks. This patch
also adds support for exclusive locks using the same algorithm as mutexes.
A new sx_init_flags() function has been added so that optional flags can be
specified to alter a given locks behavior. The flags include SX_DUPOK,
SX_NOWITNESS, SX_NOPROFILE, and SX_QUITE which are all identical in nature
to the similar flags for mutexes.
Adaptive spinning on select locks may be enabled by enabling the
ADAPTIVE_SX kernel option. Only locks initialized with the SX_ADAPTIVESPIN
flag via sx_init_flags() will adaptively spin.
The common cases for sx_slock(), sx_sunlock(), sx_xlock(), and sx_xunlock()
are now performed inline in non-debug kernels. As a result, <sys/sx.h> now
requires <sys/lock.h> to be included prior to <sys/sx.h>.
The new kernel option SX_NOINLINE can be used to disable the aforementioned
inlining in non-debug kernels.
The size of struct sx has changed, so the kernel ABI is probably greatly
disturbed.
MFC after: 1 month
Submitted by: attilio
Tested by: kris, pjd
These functions are intended to be used to drop a lock and then reacquire
it when doing an sleep such as msleep(9). Both functions accept a
'struct lock_object *' as their first parameter. The 'lc_unlock' function
returns an integer that is then passed as the second paramter to the
subsequent 'lc_lock' function. This can be used to communicate state.
For example, sx locks and rwlocks use this to indicate if the lock was
share/read locked vs exclusive/write locked.
Currently, spin mutexes and lockmgr locks do not provide working lc_lock
and lc_unlock functions.
and optimize away unused stack values. The 48 bytes that the lock_profile_object
adds to the stack evidently has a measurable performance impact on certain workloads.
- Fix missing initialization in kern_rwlock.c causing bogus times to be collected
- Move updates to the lock hash to after the lock is released for spin mutexes,
sleep mutexes, and sx locks
- Add new kernel build option LOCK_PROFILE_FAST - only update lock profiling
statistics when an acquisition is contended. This reduces the overhead of
LOCK_PROFILING to increasing system time by 20%-25% which on
"make -j8 kernel-toolchain" on a dual woodcrest is unmeasurable in terms
of wall-clock time. Contrast this to enabling lock profiling without
LOCK_PROFILE_FAST and I see a 5x-6x slowdown in wall-clock time.
- only collect timestamps when a lock is contested - this reduces the overhead
of collecting profiles from 20x to 5x
- remove unused function from subr_lock.c
- generalize cnt_hold and cnt_lock statistics to be kept for all locks
- NOTE: rwlock profiling generates invalid statistics (and most likely always has)
someone familiar with that should review
