On 04/18/2019 10:40 AM, Peter Zijlstra wrote:
> On Thu, Apr 18, 2019 at 10:08:28AM -0400, Waiman Long wrote:
>> On 04/18/2019 09:51 AM, Peter Zijlstra wrote:
>>> On Sat, Apr 13, 2019 at 01:22:57PM -0400, Waiman Long wrote:
>>>> inline void __down_read(struct rw_semaphore *sem)
>>>> {
>>>> + long count = atomic_long_fetch_add_acquire(RWSEM_READER_BIAS,
>>>> + &sem->count);
>>>> +
>>>> + if (unlikely(count & RWSEM_READ_FAILED_MASK)) {
>>>> + rwsem_down_read_failed(sem, count);
>>>> DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
>>>> } else {
>>>> rwsem_set_reader_owned(sem);
>>> *groan*, that is not provably correct. It is entirely possible to get
>>> enough fetch_add()s piled on top of one another to overflow regardless.
>>>
>>> Unlikely, yes, impossible, no.
>>>
>>> This makes me nervious as heck, I really don't want to ever have to
>>> debug something like that :-(
>> The number of fetch_add() that can pile up is limited by the number of
>> CPUs available in the system.
>> Yes, if you have a 32k processor system that have all the CPUs trying
>> to acquire the same read-lock, we will have a problem.
> Having more CPUs than that is not impossible these days.
>
Having more than 32k CPUs contending for the same cacheline will be
horribly slow.
>> Or as Linus had said that if we could have tasks kept
>> preempted right after doing the fetch_add with newly scheduled tasks
>> doing the fetch_add at the same lock again, we could have overflow with
>> less CPUs.
> That.
>
>> How about disabling preemption before fetch_all and re-enable
>> it afterward to address the latter concern?
> Performance might be an issue, look at what preempt_disable() +
> preempt_enable() generate for ARM64 for example. That's not particularly
> pretty.
That is just for the preempt kernel. Right? Thinking about it some more,
the above scenario is less likely to happen for CONFIG_PREEMPT_VOLUNTARY
kernel and the preempt_disable cost will be lower. A preempt RT kernel
is less likely to run on system with many CPUs anyway. We could make
that a conifg option as well in a follow-on patch and let the
distributors decide.
>> I have no solution for the first case, though.
> A cmpxchg() loop can fix this, but that again has performance
> implications like you mentioned a while back.
Exactly.
Cheers,
Longman
