cvs commit: src/sys/dev/fxp if_fxp.c if_fxpvar.h

[M. Warner Losh]

In message: <XFMail.20030501140515.jhb at FreeBSD.org>
            John Baldwin <jhb at FreeBSD.org> writes:
: 
: On 01-May-2003 M. Warner Losh wrote:
: > In message: <1721460000.1051803729 at aslan.btc.adaptec.com>
: >             "Justin T. Gibbs" <gibbs at scsiguy.com> writes:
: >: >> This means that all detaches must occur from a context that can
: >: >> sleep, but that shouldn't be too hard to make happen.
: >: > 
: >: > People can't hold the driver lock across bus_teardown_intr() with this
: >: > model, which does require a possibly smarter interrupt routine or
: >: > maybe a better detach that only disables interrupts then does a teardown,
: >: > then finishes shutting down the rest of the hardware along with an
: >: > interrupt handler that doesn't re-enable interrupts in the shutdown case.
: >: 
: >: All doable things for all but really broken hardware.  fxp is not broken.
: > 
: > The whole reason for the gone flag may be misunderstood here.  You can
: > easily turn off the fxp device, and there will be no more interrupts
: > from it.  However, its ISR can and will still be called from time to
: > time until the bus_teardown_intr() is complete?  Why you ask?  Because
: > of shared interrupts.  If fxp shares an interrupt with another device,
: > your ISR will execute even if you write 0 into the interrupt enable
: > register if that other device gets an interrupt between the time you
: > write to this register and the time bus_teardown_intr is called, even
: > on a single CPU machine:
: > 
: > 
: >       fxp_detach()
: > [4]   LOCK
: > [a]   write 0 to dis intr
: > [5]                           device B on same intr interrupts here
: >                               fxp_intr()
: >                               LOCK (->sleep)
: > [b]   gone = 0;
: >       UNLOCK
: > [1]                           if (gone) return;
: > [2]   bus_teardown_intr();
: > [3]   bus_teardown_intr returns
: > 
: > 
: > [1] and [2] can happen in any order, but you know both of them have
: > happened by [3].
: > 
: > The order of [a] and [b] don't really matter because fxp (or anything
: > that shares its interrupt) could generate an interrupt after the lock
: > is taken out at [4] and you'd still have a fxp_intr sleeping thread.
: > The important thing is that an interrupt[5] happens after [4].  This
: > can happen on both the single CPU case and the SMP case.
: > 
: > This might argue for blocking interrupts during a device detach.  I
: > think there might be problems with that apprach as well, although I'd
: > have to think about it a bit to be sure.
: 
: Ok, here's a question.  Why is it bad for fxp_intr() to finish while
: you are blocked in bus_teardown_intr()?  Put another way, perhaps
: fxp_detach() should do the teardown_intr() a lot sooner and postpone
: some of it's cleanups until after that retuns.  I.e.
: 
:         FXP_LOCK()
:         disable_interrupts_in_hardware()
:         FXP_UNLOCK()
:         bus_teardown_intr()
:         FXP_LOCK()
:         do_rest_of_detach()

This means you don't have to call the 'is my hardware there' routine
in the isr.  That's what gone buys you.  I'm not sure that's
necessarily a big thing...  This does imply that we may want two ISRs,
one for hot-swappable hardware and one for not, at least for the high
performance drivers.  Lower performance ones can elect to not care
about the slight extra overhead of checking.

Warner