We found that a process with 10 thousnads threads has been encountered
a regression problem from Linux-v4.14 to Linux-v5.4. It is a kind of
workload which will concurrently allocate lots of memory in different
threads sometimes. In this case, we will see the down_read_trylock()
with a high hotspot. Therefore, we suppose that rwsem has a regression
at least since Linux-v5.4. In order to easily debug this problem, we
write a simply benchmark to create the similar situation lile the
following.
```c++
#include <sys/mman.h>
#include <sys/time.h>
#include <sys/resource.h>
#include <sched.h>
#include <cstdio>
#include <cassert>
#include <thread>
#include <vector>
#include <chrono>
volatile int mutex;
void trigger(int cpu, char* ptr, std::size_t sz)
{
cpu_set_t set;
CPU_ZERO(&set);
CPU_SET(cpu, &set);
assert(pthread_setaffinity_np(pthread_self(), sizeof(set), &set) == 0);
while (mutex);
for (std::size_t i = 0; i < sz; i += 4096) {
*ptr = '\0';
ptr += 4096;
}
}
int main(int argc, char* argv[])
{
std::size_t sz = 100;
if (argc > 1)
sz = atoi(argv[1]);
auto nproc = std::thread::hardware_concurrency();
std::vector<std::thread> thr;
sz <<= 30;
auto* ptr = mmap(nullptr, sz, PROT_READ | PROT_WRITE, MAP_ANON |
MAP_PRIVATE, -1, 0);
assert(ptr != MAP_FAILED);
char* cptr = static_cast<char*>(ptr);
auto run = sz / nproc;
run = (run >> 12) << 12;
mutex = 1;
for (auto i = 0U; i < nproc; ++i) {
thr.emplace_back(std::thread([i, cptr, run]() { trigger(i, cptr, run); }));
cptr += run;
}
rusage usage_start;
getrusage(RUSAGE_SELF, &usage_start);
auto start = std::chrono::system_clock::now();
mutex = 0;
for (auto& t : thr)
t.join();
rusage usage_end;
getrusage(RUSAGE_SELF, &usage_end);
auto end = std::chrono::system_clock::now();
timeval utime;
timeval stime;
timersub(&usage_end.ru_utime, &usage_start.ru_utime, &utime);
timersub(&usage_end.ru_stime, &usage_start.ru_stime, &stime);
printf("usr: %ld.%06ld\n", utime.tv_sec, utime.tv_usec);
printf("sys: %ld.%06ld\n", stime.tv_sec, stime.tv_usec);
printf("real: %lu\n",
std::chrono::duration_cast<std::chrono::milliseconds>(end -
start).count());
return 0;
}
```
The functionality of above program is simply which creates `nproc`
threads and each of them are trying to touch memory (trigger page
fault) on different CPU. Then we will see the similar profile by
`perf top`.
25.55% [kernel] [k] down_read_trylock
14.78% [kernel] [k] handle_mm_fault
13.45% [kernel] [k] up_read
8.61% [kernel] [k] clear_page_erms
3.89% [kernel] [k] __do_page_fault
The highest hot instruction, which accounts for about 92%, in
down_read_trylock() is cmpxchg like the following.
91.89 │ lock cmpxchg %rdx,(%rdi)
Sice the problem is found by migrating from Linux-v4.14 to Linux-v5.4,
so we easily found that the commit ddb20d1d3a ("locking/rwsem: Optimize
down_read_trylock()") caused the regression. The reason is that the
commit assumes the rwsem is not contended at all. But it is not always
true for mmap lock which could be contended with thousands threads.
So most threads almost need to run at least 2 times of "cmpxchg" to
acquire the lock. The overhead of atomic operation is higher than
non-atomic instructions, which caused the regression.
By using the above benchmark, the real executing time on a x86-64 system
before and after the patch were:
Before Patch After Patch
# of Threads real real reduced by
------------ ------ ------ ----------
1 65,373 65,206 ~0.0%
4 15,467 15,378 ~0.5%
40 6,214 5,528 ~11.0%
For the uncontended case, the new down_read_trylock() is the same as
before. For the contended cases, the new down_read_trylock() is faster
than before. The more contended, the more fast.
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Link: https://lore.kernel.org/r/20211118094455.9068-1-songmuchun@bytedance.com
There are some inconsistency in the way that the handoff bit is being
handled in readers and writers that lead to a race condition.
Firstly, when a queue head writer set the handoff bit, it will clear
it when the writer is being killed or interrupted on its way out
without acquiring the lock. That is not the case for a queue head
reader. The handoff bit will simply be inherited by the next waiter.
Secondly, in the out_nolock path of rwsem_down_read_slowpath(), both
the waiter and handoff bits are cleared if the wait queue becomes
empty. For rwsem_down_write_slowpath(), however, the handoff bit is
not checked and cleared if the wait queue is empty. This can
potentially make the handoff bit set with empty wait queue.
Worse, the situation in rwsem_down_write_slowpath() relies on wstate,
a variable set outside of the critical section containing the ->count
manipulation, this leads to race condition where RWSEM_FLAG_HANDOFF
can be double subtracted, corrupting ->count.
To make the handoff bit handling more consistent and robust, extract
out handoff bit clearing code into the new rwsem_del_waiter() helper
function. Also, completely eradicate wstate; always evaluate
everything inside the same critical section.
The common function will only use atomic_long_andnot() to clear bits
when the wait queue is empty to avoid possible race condition. If the
first waiter with handoff bit set is killed or interrupted to exit the
slowpath without acquiring the lock, the next waiter will inherit the
handoff bit.
While at it, simplify the trylock for loop in
rwsem_down_write_slowpath() to make it easier to read.
Fixes: 4f23dbc1e6 ("locking/rwsem: Implement lock handoff to prevent lock starvation")
Reported-by: Zhenhua Ma <mazhenhua@xiaomi.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20211116012912.723980-1-longman@redhat.com
Guard the regular sleeping lock specific functionality, which is used for
rtmutex on non-RT enabled kernels and for mutex, rtmutex and semaphores on
RT enabled kernels so the code can be reused for the RT specific
implementation of spinlocks and rwlocks in a different compilation unit.
No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20210815211303.311535693@linutronix.de
The RT specific R/W semaphore implementation used to restrict the number of
readers to one, because a writer cannot block on multiple readers and
inherit its priority or budget.
The single reader restricting was painful in various ways:
- Performance bottleneck for multi-threaded applications in the page fault
path (mmap sem)
- Progress blocker for drivers which are carefully crafted to avoid the
potential reader/writer deadlock in mainline.
The analysis of the writer code paths shows that properly written RT tasks
should not take them. Syscalls like mmap(), file access which take mmap sem
write locked have unbound latencies, which are completely unrelated to mmap
sem. Other R/W sem users like graphics drivers are not suitable for RT tasks
either.
So there is little risk to hurt RT tasks when the RT rwsem implementation is
done in the following way:
- Allow concurrent readers
- Make writers block until the last reader left the critical section. This
blocking is not subject to priority/budget inheritance.
- Readers blocked on a writer inherit their priority/budget in the normal
way.
There is a drawback with this scheme: R/W semaphores become writer unfair
though the applications which have triggered writer starvation (mostly on
mmap_sem) in the past are not really the typical workloads running on a RT
system. So while it's unlikely to hit writer starvation, it's possible. If
there are unexpected workloads on RT systems triggering it, the problem
has to be revisited.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
Reader optimistic spinning is helpful when the reader critical section
is short and there aren't that many readers around. It also improves
the chance that a reader can get the lock as writer optimistic spinning
disproportionally favors writers much more than readers.
Since commit d3681e269f ("locking/rwsem: Wake up almost all readers
in wait queue"), all the waiting readers are woken up so that they can
all get the read lock and run in parallel. When the number of contending
readers is large, allowing reader optimistic spinning will likely cause
reader fragmentation where multiple smaller groups of readers can get
the read lock in a sequential manner separated by writers. That reduces
reader parallelism.
One possible way to address that drawback is to limit the number of
readers (preferably one) that can do optimistic spinning. These readers
act as representatives of all the waiting readers in the wait queue as
they will wake up all those waiting readers once they get the lock.
Alternatively, as reader optimistic lock stealing has already enhanced
fairness to readers, it may be easier to just remove reader optimistic
spinning and simplifying the optimistic spinning code as a result.
Performance measurements (locking throughput kops/s) using a locking
microbenchmark with 50/50 reader/writer distribution and turbo-boost
disabled was done on a 2-socket Cascade Lake system (48-core 96-thread)
to see the impacts of these changes:
1) Vanilla - 5.10-rc3 kernel
2) Before - 5.10-rc3 kernel with previous patches in this series
2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch
3) no-rspin - 5.10-rc3 kernel with reader spinning disabled
# of threads CS Load Vanilla Before limit-rspin no-rspin
------------ ------- ------- ------ ----------- --------
2 1 5,185 5,662 5,214 5,077
4 1 5,107 4,983 5,188 4,760
8 1 4,782 4,564 4,720 4,628
16 1 4,680 4,053 4,567 3,402
32 1 4,299 1,115 1,118 1,098
64 1 3,218 983 1,001 957
96 1 1,938 944 957 930
2 20 2,008 2,128 2,264 1,665
4 20 1,390 1,033 1,046 1,101
8 20 1,472 1,155 1,098 1,213
16 20 1,332 1,077 1,089 1,122
32 20 967 914 917 980
64 20 787 874 891 858
96 20 730 836 847 844
2 100 372 356 360 355
4 100 492 425 434 392
8 100 533 537 529 538
16 100 548 572 568 598
32 100 499 520 527 537
64 100 466 517 526 512
96 100 406 497 506 509
The column "CS Load" represents the number of pause instructions issued
in the locking critical section. A CS load of 1 is extremely short and
is not likey in real situations. A load of 20 (moderate) and 100 (long)
are more realistic.
It can be seen that the previous patches in this series have reduced
performance in general except in highly contended cases with moderate
or long critical sections that performance improves a bit. This change
is mostly caused by the "Prevent potential lock starvation" patch that
reduce reader optimistic spinning and hence reduce reader fragmentation.
The patch that further limit reader optimistic spinning doesn't seem to
have too much impact on overall performance as shown in the benchmark
data.
The patch that disables reader optimistic spinning shows reduced
performance at lightly loaded cases, but comparable or slightly better
performance on with heavier contention.
This patch just removes reader optimistic spinning for now. As readers
are not going to do optimistic spinning anymore, we don't need to
consider if the OSQ is empty or not when doing lock stealing.
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Davidlohr Bueso <dbueso@suse.de>
Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
If the optimistic spinning queue is empty and the rwsem does not have
the handoff or write-lock bits set, it is actually not necessary to
call rwsem_optimistic_spin() to spin on it. Instead, it can steal the
lock directly as its reader bias is in the count already. If it is
the first reader in this state, it will try to wake up other readers
in the wait queue.
With this patch applied, the following were the lock event counts
after rebooting a 2-socket system and a "make -j96" kernel rebuild.
rwsem_opt_rlock=4437
rwsem_rlock=29
rwsem_rlock_steal=19
So lock stealing represents about 0.4% of all the read locks acquired
in the slow path.
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Davidlohr Bueso <dbueso@suse.de>
Link: https://lkml.kernel.org/r/20201121041416.12285-4-longman@redhat.com
The lock handoff bit is added in commit 4f23dbc1e6 ("locking/rwsem:
Implement lock handoff to prevent lock starvation") to avoid lock
starvation. However, allowing readers to do optimistic spinning does
introduce an unlikely scenario where lock starvation can happen.
The lock handoff bit may only be set when a waiter is being woken up.
In the case of reader unlock, wakeup happens only when the reader count
reaches 0. If there is a continuous stream of incoming readers acquiring
read lock via optimistic spinning, it is possible that the reader count
may never reach 0 and so the handoff bit will never be asserted.
One way to prevent this scenario from happening is to disallow optimistic
spinning if the rwsem is currently owned by readers. If the previous
or current owner is a writer, optimistic spinning will be allowed.
If the previous owner is a reader but the reader count has reached 0
before, a wakeup should have been issued. So the handoff mechanism
will be kicked in to prevent lock starvation. As a result, it should
be OK to do optimistic spinning in this case.
This patch may have some impact on reader performance as it reduces
reader optimistic spinning especially if the lock critical sections
are short the number of contending readers are small.
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Davidlohr Bueso <dbueso@suse.de>
Link: https://lkml.kernel.org/r/20201121041416.12285-3-longman@redhat.com
In preparation for converting exec_update_mutex to a rwsem so that
multiple readers can execute in parallel and not deadlock, add
down_read_interruptible. This is needed for perf_event_open to be
converted (with no semantic changes) from working on a mutex to
wroking on a rwsem.
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/87k0tybqfy.fsf@x220.int.ebiederm.org
In preparation for converting exec_update_mutex to a rwsem so that
multiple readers can execute in parallel and not deadlock, add
down_read_killable_nested. This is needed so that kcmp_lock
can be converted from working on a mutexes to working on rw_semaphores.
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/87o8jabqh3.fsf@x220.int.ebiederm.org
Extend lockdep to validate lock wait-type context.
The current wait-types are:
LD_WAIT_FREE, /* wait free, rcu etc.. */
LD_WAIT_SPIN, /* spin loops, raw_spinlock_t etc.. */
LD_WAIT_CONFIG, /* CONFIG_PREEMPT_LOCK, spinlock_t etc.. */
LD_WAIT_SLEEP, /* sleeping locks, mutex_t etc.. */
Where lockdep validates that the current lock (the one being acquired)
fits in the current wait-context (as generated by the held stack).
This ensures that there is no attempt to acquire mutexes while holding
spinlocks, to acquire spinlocks while holding raw_spinlocks and so on. In
other words, its a more fancy might_sleep().
Obviously RCU made the entire ordeal more complex than a simple single
value test because RCU can be acquired in (pretty much) any context and
while it presents a context to nested locks it is not the same as it
got acquired in.
Therefore its necessary to split the wait_type into two values, one
representing the acquire (outer) and one representing the nested context
(inner). For most 'normal' locks these two are the same.
[ To make static initialization easier we have the rule that:
.outer == INV means .outer == .inner; because INV == 0. ]
It further means that its required to find the minimal .inner of the held
stack to compare against the outer of the new lock; because while 'normal'
RCU presents a CONFIG type to nested locks, if it is taken while already
holding a SPIN type it obviously doesn't relax the rules.
Below is an example output generated by the trivial test code:
raw_spin_lock(&foo);
spin_lock(&bar);
spin_unlock(&bar);
raw_spin_unlock(&foo);
[ BUG: Invalid wait context ]
-----------------------------
swapper/0/1 is trying to lock:
ffffc90000013f20 (&bar){....}-{3:3}, at: kernel_init+0xdb/0x187
other info that might help us debug this:
1 lock held by swapper/0/1:
#0: ffffc90000013ee0 (&foo){+.+.}-{2:2}, at: kernel_init+0xd1/0x187
The way to read it is to look at the new -{n,m} part in the lock
description; -{3:3} for the attempted lock, and try and match that up to
the held locks, which in this case is the one: -{2,2}.
This tells that the acquiring lock requires a more relaxed environment than
presented by the lock stack.
Currently only the normal locks and RCU are converted, the rest of the
lockdep users defaults to .inner = INV which is ignored. More conversions
can be done when desired.
The check for spinlock_t nesting is not enabled by default. It's a separate
config option for now as there are known problems which are currently
addressed. The config option allows to identify these problems and to
verify that the solutions found are indeed solving them.
The config switch will be removed and the checks will permanently enabled
once the vast majority of issues has been addressed.
[ bigeasy: Move LD_WAIT_FREE,… out of CONFIG_LOCKDEP to avoid compile
failure with CONFIG_DEBUG_SPINLOCK + !CONFIG_LOCKDEP]
[ tglx: Add the config option ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200321113242.427089655@linutronix.de
