Updated rusage patch
Jeff Roberson
jroberson at chesapeake.net
Fri Jun 1 09:28:04 UTC 2007
On Fri, 1 Jun 2007, Bruce Evans wrote:
> On Fri, 1 Jun 2007, Bruce Evans wrote:
>
>> Hmm, this is confusing. Normal locking is not used for thread-local
>> fields. Instead, a side effect of spinlocking is used:
>> mtx_lock_spin(&any) in non-interrupt code has the side effect of
>> masking interrupts on the current CPU, so statclock() can access
>> anything on the current CPU without acquiring any locks, just like
>> an interrupt handler on a UP system. This is used for td_*ticks.
>> It doesn't work for ru_*rss since since exit1() doesn't hold a
>> spinlock when copying td_ru. The sched_locking of ru_*rss in
>> statclock() doesn't help -- I think it now does nothing except
>> waste time, since these fields are now thread-local so they need
>> the same locking as td_*ticks, which is null in statclock() but
>> non-null elsewhere.
>>
>> Related bugs:
>> - td_[usip]ticks are still under (j) (sched_lock) in proc.h.
>> - td_(uu,us}ticks have always (?) been under (k) (thread-local). That
>> is more correct than (j), but they are updated by an interrupt handler
>> and seem to be accessed without disabling interrupts elsewhere.
>
> Oops, it's more confusing than that. It is not a bug for td_[usip]ticks
> to be under sched_lock, since they must be under a more specific lock
> than `any' for when they are reset in ruxagg(). That lock has the dual
> purpose of locking out interrupts as a side effect and locking out other
> threads explicitly.
>
> Please add a lock assertion in ruxagg() and friends that the relevant
> lock is held.
I have such an assert in a threadlock.diff that I will post tomorrow. I
could add it in before as well, but I suspect threadlock is going in next
week and there are different locks involved there anyway.
Jeff
>
> Bruce
>
More information about the freebsd-arch
mailing list