Date: Tue, 19 Jan 2010 10:53:23 +0100 From: Attilio Rao <attilio@freebsd.org> To: Jeff Roberson <jroberson@jroberson.net> Cc: freebsd-current@freebsd.org, Kohji Okuno <okuno.kohji@jp.panasonic.com> Subject: Re: Bug about sched_4bsd? Message-ID: <3bbf2fe11001190153s1d42a020pecb343993b7971a2@mail.gmail.com> In-Reply-To: <3bbf2fe11001190152k15c24f70k876762817bf522c1@mail.gmail.com> References: <3bbf2fe11001171858o4568fe38l9b2db54ec9856b50@mail.gmail.com> <20100118.155352.59640143160034670.okuno.kohji@jp.panasonic.com> <3bbf2fe11001172306m69ff6544i3aaf05e2540136e1@mail.gmail.com> <20100119.103858.29593248145858473.okuno.kohji@jp.panasonic.com> <alpine.BSF.2.00.1001181544130.1027@desktop> <3bbf2fe11001190152k15c24f70k876762817bf522c1@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
2010/1/19 Attilio Rao <attilio@freebsd.org>: > 2010/1/19 Jeff Roberson <jroberson@jroberson.net>: >> On Tue, 19 Jan 2010, Kohji Okuno wrote: >> >>> Hello, Attilio, >>> >>> I think setpriority() can set priority to sleeping threads. >>> Is it really safe? >> >> I agree, in this code path maybe_resched is not properly locking curthre= ad. >> =C2=A0curthread will be sched_lock and the sleeping thread will be a sle= epq lock. >> =C2=A0I believe that since &sched_lock is ordered after container locks = it would >> be sufficient to compare the two td_lock pointers and acquire sched_lock= if >> they are not equal. =C2=A0Someone should look at other maybe_resched cal= lers or >> add an assert that curthread->td_lock is always &sched_lock in this >> function. > > I'm not sure I understand you well here, but I generally don't agree, > if we speak about the current code plus the patch I posted. > Without the patch, there is a general problem of maybe_preempt() > because sched_switch() will handle TDF_NEEDRESCHED just in racy ways > (not ensuring atomicity of td_lock operations for sleeping threads). > That's, however, still not specific to maybe_preempt() only. However: > * If you make a problem about the callers of maybe_resched() I agree. > The callers should assert for sched_lock to be in place. But that is > not a general problem of maybe_resched(), it is on the callers > ballpark > * If you make a problem about the locking itself, the patch IMHO > should fix it or there is still something I can't see. s/maybe_preempt/maybe_resched, of course :( Attilio --=20 Peace can only be achieved by understanding - A. Einstein
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3bbf2fe11001190153s1d42a020pecb343993b7971a2>