Previous limits were chosen when locking primitives had spurious lock
accesses.
Flipping the starting point to 1 (or rather 2 as the first call shifts it)
provides a modest win when mild contention is seen while not hurting worse
cases. Tested on a bunch of one, two and four socket old and new systems
(Westmere, Skylake, Threadreaper and others) by doing concurrent page faults,
buildkernel/buildworld and other stuff (although not all systems got all the
tests).
Another thing is the upper limit. It is semi-arbitrarily chosen as it was
getting out of hand for slightly less small systems (e.g. a 128-thread one).
Note that backoff is fundamentally a speculative bandaid and this change just
makes it fit a little bit better. It remains completely oblivious to the
hardware topology or the contention pattern. This is being experimented with.
Mainly focus on files that use BSD 2-Clause license, however the tool I
was using misidentified many licenses so this was mostly a manual - error
prone - task.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
Previous implementation would use a random factor to spread readers and
reduce chances of starvation. This visibly reduces effectiveness of the
mechanism.
Switch to the more traditional exponential variant. Try to limit starvation
by imposing an upper limit of spins after which spinning is half of what
other threads get. Note the mechanism is turned off by default.
Reviewed by: kib (previous version)
All current spinning loops retry an atomic op the first chance they get,
which leads to performance degradation under load.
One classic solution to the problem consists of delaying the test to an
extent. This implementation has a trivial linear increment and a random
factor for each attempt.
For simplicity, this first thouch implementation only modifies spinning
loops where the lock owner is running. spin mutexes and thread lock were
not modified.
Current parameters are autotuned on boot based on mp_cpus.
Autotune factors are very conservative and are subject to change later.
Reviewed by: kib, jhb
Tested by: pho
MFC after: 1 week
A _NEW flag passed to _init_flags() to avoid check for double-init.
Differential Revision: https://reviews.freebsd.org/D1208
Reviewed by: jhb, wblock
MFC after: 1 Month
provided by Isilon.
- Add an rm_assert() supporting various lock assertions similar to other
locking primitives. Because rmlocks track readers the assertions are
always fully accurate unlike rw_assert() and sx_assert().
- Flesh out the lock class methods for rmlocks to support sleeping via
condvars and rm_sleep() (but only while holding write locks), rmlock
details in 'show lock' in DDB, and the lc_owner method used by
dtrace.
- Add an internal destroyed cookie so that API functions can assert
that an rmlock is not destroyed.
- Make use of rm_assert() to add various assertions to the API (e.g.
to assert locks are held when an unlock routine is called).
- Give RM_SLEEPABLE locks their own lock class and always use the
rmlock's own lock_object with WITNESS.
- Use THREAD_NO_SLEEPING() / THREAD_SLEEPING_OK() to disallow sleeping
while holding a read lock on an rmlock.
Submitted by: andre
Obtained from: EMC/Isilon
- Implement a function to ensure that all preempted threads have switched
back out at least once. Use this to make sure there are no stale
references to the old ktr_buf or the lock profiling buffers before
updating them.
Reviewed by: marius (sparc64 parts), attilio (earlier patch)
Sponsored by: EMC / Isilon Storage Division
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)
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.
sbuf_new_for_sysctl(9). This allows using an sbuf with a SYSCTL_OUT
drain for extremely large amounts of data where the caller knows that
appropriate references are held, and sleeping is not an issue.
Inspired by: rwatson
Add a drain function for struct sysctl_req, and use it for a variety
of handlers, some of which had to do awkward things to get a large
enough SBUF_FIXEDLEN buffer.
Note that some sysctl handlers were explicitly outputting a trailing
NUL byte. This behaviour was preserved, though it should not be
necessary.
Reviewed by: phk (original patch)
unexpected things in copyout(9) and so wiring the user buffer is not
sufficient to perform a copyout(9) while holding a random mutex.
Requested by: nwhitehorn
handlers, some of which had to do awkward things to get a large enough
FIXEDLEN buffer.
Note that some sysctl handlers were explicitly outputting a trailing NUL
byte. This behaviour was preserved, though it should not be necessary.
Reviewed by: phk
guarantee that all cpus have acknowledged the cleared enable int by
scheduling the resetting thread on each cpu in succession. Since all
lock profiling happens within a critical section this guarantees that
all cpus have left lock profiling before we clear the datastructures.
- Assert that the per-thread queue of locks lock profiling is aware of
is clear on thread exit. There were several cases where this was not
true that slows lock profiling and leaks information.
- Remove all objects from all lists before clearing any per-cpu
information in reset. Lock profiling objects can migrate between
per-cpu caches and previously these migrated objects could be zero'd
before they'd been removed
Discussed with: attilio
Sponsored by: Nokia
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
done in consumers code: using locks properties is much more appropriate.
Fix current code doing these bogus checks.
Note: Really, callout are not usable by all !(LC_SPINLOCK | LC_SLEEPABLE)
primitives like rmlocks doesn't implement the generic lock layer
functions, but they can be equipped for this, so the check is still
valid.
Tested by: matteo, kris (earlier version)
Reviewed by: jhb
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
lock optimized for almost exclusive reader access. (see also rmlock.9)
TODO:
Convert to per cpu variables linkerset as soon as it is available.
Optimize UP (single processor) case.
per-primitive macros like MTX_NOPROFILE, SX_NOPROFILE or RW_NOPROFILE) is
not really honoured. In particular lock_profile_obtain_lock_failure() and
lock_profile_obtain_lock_success() are naked respect this flag.
The bug leads to locks marked with no-profiling to be profiled as well.
In the case of the clock_lock, used by the timer i8254 this leads to
unpredictable behaviour both on amd64 and ia32 (double faults panic,
sudden reboots, etc.). The amd64 clock_lock is also not marked as
not profilable as it should be.
Fix these bugs adding proper checks in the lock profiling code and at
clock_lock initialization time.
i8254 bug pointed out by: kris
Tested by: matteo, Giuseppe Cocomazzi <sbudella at libero dot it>
Approved by: jeff (mentor)
Approved by: re
- 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
if waittime was zero (the lock was uncontested) l->lpo_waittime
in the hash table would not get initialized.
Inspection prompted by questions from: Attilio Rao
wait (time waited to acquire) and hold times for *all* kernel locks. If
the architecture has a system synchronized TSC, the profiling code will
use that - thereby minimizing profiling overhead. Large chunks of profiling
code have been moved out of line, the overhead measured on the T1 for when
it is compiled in but not enabled is < 1%.
Approved by: scottl (standing in for mentor rwatson)
Reviewed by: des and jhb
implementation is by no means perfect as far as some of the algorithms
that it uses and the fact that it is missing some functionality (try
locks and upgrades/downgrades are not there yet), however it does seem
to work in my local testing. There is more detail in the comments in the
code, but the short version follows.
A reader/writer lock is very much like a regular mutex: it cannot be held
across a voluntary sleep; it can be acquired in an interrupt thread; if
the lock is held by a writer then the priority of any threads that block
on the lock will be lent to the owner; the simple case lock operations all
are done in a single atomic op. It also shares some similiarities
with sx locks: it supports reader/writer semantics (multiple readers,
but single writers); readers are allowed to recurse, but writers are not.
We can extend this implementation further by either improving algorithms
or adding new functionality, but this should at least give us a base to
work with now.
Reviewed by: arch (in theory)
Tested on: i386 (4 cpu box with a kernel module that used 4 threads
that randomly chose between read locks and write locks
that ran w/o panicing for over a day solid. It usually
panic'd within a few seconds when there were bugs during
testing. :) The kernel module source is available on
request.)
lock_obj objects:
- Add new lock_init() and lock_destroy() functions to setup and teardown
lock_object objects including KTR logging and registering with WITNESS.
- Move all the handling of LO_INITIALIZED out of witness and the various
lock init functions into lock_init() and lock_destroy().
- Remove the constants for static indices into the lock_classes[] array
and change the code outside of subr_lock.c to use LOCK_CLASS to compare
against a known lock class.
- Move the 'show lock' ddb function and lock_classes[] array out of
kern_mutex.c over to subr_lock.c.
