Date: Tue, 26 Feb 2002 10:02:34 -0500 (EST) From: John Baldwin <jhb@FreeBSD.org> To: Bruce Evans <bde@zeta.org.au> Cc: Terry Lambert <tlambert2@mindspring.com>, Alfred Perlstein <alfred@FreeBSD.org>, Bosko Milekic <bmilekic@unixdaemons.com>, Seigo Tanimura <tanimura@r.dl.itc.u-tokyo.ac.jp>, current@FreeBSD.org, Matthew Dillon <dillon@apollo.backplane.com> Subject: Re: Patch for critical_enter()/critical_exit() & interrupt assem Message-ID: <XFMail.020226100234.jhb@FreeBSD.org> In-Reply-To: <20020226185416.P43865-100000@gamplex.bde.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 26-Feb-02 Bruce Evans wrote: > On Tue, 26 Feb 2002, John Baldwin wrote: > >> The critical section stuff currently in current is part of the original >> preemption patches I wrote at Usenix last year. They aren't in the tree >> because they aren't stable yet. We still have problems on the alpha and >> problems with IPI deadlocks on SMP machines that need to be worked out. >> Since >> this API is still very much in flux I'd prefer to keep it simple for now and >> not make the code overly complex with optimizations until after we have >> settled >> on the design. > > The change is mostly part of the design change needed to un-break fast > interrupt handlers. critical_enter() mask not mask hardware > interrupts for the same reason than splhigh() must not mask them: > there is too much MI code that thinks it is critical and short but > really isn't (e.g., everything under spinlocks in -current). Changing > cpu_critical_enter() to a null version to prevent spinlocks masking > interrupts doesn't work very well because it is used for other things > that really do need to mask interrupts. Having 2 levels for > cpu_critical_enter() (on that masks normal interrupts and one that > masks fast interrupts) would work but I think there are already too > many levels of critical_enter()s. I think having the sparc-like API of intr_disable() / intr_restore() that we used to use to fix the places that assume too many implementation details of cpu_critical_foo() would be a good first step so that cpu_critical_foo() can be relegated back to supporting non-preemption (and no other spinlocks) and nothing else. > Bruce -- John Baldwin <jhb@FreeBSD.org> <>< http://www.FreeBSD.org/~jhb/ "Power Users Use the Power to Serve!" - http://www.FreeBSD.org/ To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?XFMail.020226100234.jhb>