Writers waiting on readers to finish can set the RW_LOCK_WRITE_SPINNER
bit. This prevents most new readers from coming on. However, the last
reader to unlock also clears the bit which means new readers can sneak
in and the cycle starts over.
Change the code to keep the bit after last unlock.
Note that starvation potential is still there: no matter how many write
spinners are there, there is one bit. After the writer unlocks, the lock
is free to get raided by readers again. It is good enough for the time
being.
The real fix would include counting writers.
This runs into a caveat: the writer which set the bit may now be preempted.
In order to get rid of the problem all attempts to set the bit are preceeded
with critical_enter.
The bit gets cleared when the thread which set it goes to sleep. This way
an invariant holds that if the bit is set, someone is actively spinning and
will grab the lock soon. In particular this means that readers which find
the lock in this transient state can safely spin until the lock finds itself
an owner (i.e. they don't need to block nor speculate how long to spin
speculatively).
Tested by: pho
Now that 10 years have passed since the original limit of 10000 was
committed, bump it a little bit.
Spinning waiting for writers is semi-informed in the sense that we always
know if the owner is running and base the decision to spin on that.
However, no such information is provided for read-locking. In particular
this means that it is possible for a write-spinner to completely waste cpu
time waiting for the lock to be released, while the reader holding it was
preempted and is now waiting for the spinner to go off cpu.
Nonetheless, in majority of cases it is an improvement to spin instead of
instantly giving up and going to sleep.
The current approach is pretty simple: snatch the number of current readers
and performs that many pauses before checking again. The total number of
pauses to execute is limited to 10k. If the lock is still not free by
that time, go to sleep.
Given the previously noted problem of not knowing whether spinning makes
any sense to begin with the new limit has to remain rather conservative.
But at the very least it should also be related to the machine. Waiting
for writers uses parameters selected based on the number of activated
hardware threads. The upper limit of pause instructions to be executed
in-between re-reads of the lock is typically 16384 or 32678. It was
selected as the limit of total spins. The lower bound is set to
already present 10000 as to not change it for smaller machines.
Bumping the limit reduces system time by few % during benchmarks like
buildworld, buildkernel and others. Tested on 2 and 4 socket machines
(Broadwell, Skylake).
Figuring out how to make a more informed decision while not pessimizing
the fast path is left as an exercise for the reader.
The slow path is always taken when lockstat is enabled. This induces
rdtsc (or other) calls to get the cycle count even when there was no
contention.
Still go to the slow path to not mess with the fast path, but avoid
the heavy lifting unless necessary.
This reduces sys and real time during -j 80 buildkernel:
before: 3651.84s user 1105.59s system 5394% cpu 1:28.18 total
after: 3685.99s user 975.74s system 5450% cpu 1:25.53 total
disabled: 3697.96s user 411.13s system 5261% cpu 1:18.10 total
So note this is still a significant hit.
LOCK_PROFILING results are not affected.
If there were exactly rowner_retries/asx_retries (by default: 10) transitions
between read and write state and the waiters still did not get the lock, the
next owner -> reader transition would result in the code correctly falling
back to turnstile/sleepq where it would incorrectly think it was waiting
for a writer and decide to leave turnstile/sleepq to loop back. From this
point it would take ts/sq trips until the lock gets released.
The bug sometimes manifested itself in stalls during -j 128 package builds.
Refactor the code to fix the bug, while here remove some of the gratituous
differences between rw and sx locks.
With the option used to compile the kernel both sx and rw shared ops would
always go to the slow path which added avoidable overhead even when the
facility is disabled.
Furthermore the increased time spent doing uncontested shared lock acquire
would be bogusly added to total wait time, somewhat skewing the results.
Restore old behaviour of going there only when profiling is enabled.
This change is a no-op for kernels without LOCK_PROFILING (which is the
default).
In both rw and sx locks we always go to sleep if the lock owner is not
running.
We do spin for some time if the lock is read-locked.
However, if we decide to go to sleep due to the lock owner being off cpu
and after sleepq/turnstile gets acquired the lock is read-locked, we should
fallback to the aforementioned wait.
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.
When waiters/writer spinner flags are set no new readers can show up unless
they already have a different rw rock read locked. The change in r326195 failed
to take that into account - in presence of new readers it would spin until
they all drain, which would be lead to trouble if e.g. they go off cpu and
can get scheduled because of this thread.
Reported by: pho
In order to go to sleep threads set waiter flags, but that can spuriously
fail e.g. when a new reader arrives. Instead of unlocking everything and
looping back, re-evaluate the new state while still holding the lock necessary
to go to sleep.
The pair is of use only in debug or LOCKPROF kernels, but was passed (zeroed)
for many locks even in production kernels.
While here whack the tid argument from wlock hard and xlock hard.
There is no kbi change of any sort - "external" primitives still accept the
pair.
Previous code would always spin once before checking the lock. But a lock
with e.g. 6 readers is not going to become free in the duration of once spin
even if they start draining immediately.
Conservatively perform one for each reader.
Note that the total number of allowed spins is still extremely small and is
subject to change later.
MFC after: 1 week
spin first instant of instantly re-readoing and don't re-read after
spinning is finished - the state is already known.
Note the code is subject to significant changes later.
MFC after: 1 week
Most of the lock slowpaths assert that the calling thread isn't an idle
thread. However, this may not be true if the system has panicked, and in
some cases the assertion appears before a SCHEDULER_STOPPED() check.
MFC after: 3 days
Sponsored by: Dell EMC Isilon
Unclear how, but the locking routine for mutexes was using the *release*
barrier instead of acquire. This must have been either a copy-pasto or bad
completion.
Going through other uses of atomics shows no barriers in:
- upgrade routines (addressed in this patch)
- sections protected with turnstile locks - this should be fine as necessary
barriers are in the worst case provided by turnstile unlock
I would like to thank Mark Millard and andreast@ for reporting the problem and
testing previous patches before the issue got identified.
ps.
.-'---`-.
,' `.
| \
| \
\ _ \
,\ _ ,'-,/-)\
( * \ \,' ,' ,'-)
`._,) -',-')
\/ ''/
) / /
/ ,'-'
Hardware provided by: IBM LTC
Since fcmpset can fail without lock contention e.g. on arm, it was possible
to get spurious failures when the caller was expecting the primitive to succeed.
Reported by: mmel
It is only needed if the LOCK_PROFILING is enabled. It has to always check if
the lock is about to be released which requires an avoidable read if the option
is not specified..
They all fallback to the slow path if necessary and the check is there.
This means a panicked kernel executing code from modules will be able to
succeed doing actual lock/unlock, but this was already the case for core code
which has said primitives inlined.
Update comments to note these functions are reachable if lockstat is
enabled.
Check if the lock has any bits set before attempting unlock, which saves
an unnecessary atomic operation.
This improves singlethreaded throughput on my test machine from ~247 mln
ops/s to ~328 mln.
It is mostly about avoiding the setup cost of lockstat.
Reviewed by: jhb (previous version)
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)
When a relevant lockstat probe is enabled the fallback primitive is called with
a constant signifying a free lock. This works fine for typical cases but breaks
with recursion, since it checks if the passed value is that of the executing
thread.
Read the value if necessary.
See r313275 for details.
One difference here is that recursion handling was removed from the fallback
routine. As it is it was never supposed to see a recursed lock in the first
place. Future changes will move it out of inline variants, but right now
there is no easy to way to test if the lock is recursed without reading
additional words.