Date: Tue, 25 Apr 1995 16:50:50 +1000 From: Bruce Evans <bde@zeta.org.au> To: bde@zeta.org.au, matt@lkg.dec.com Cc: hackers@FreeBSD.org, julian@ref.tfs.com, phk@ref.tfs.com Subject: Re: [EISA] related matters Message-ID: <199504250650.QAA19925@godzilla.zeta.org.au>
next in thread | raw e-mail | index | archive | help
>> >What we really should have is a TEXT_SET(device_probe...) and kill config... >> >> No linker magic is required or good. Drivers should register their >> interrupt masking requirements as late as possible, preferably not >> until open() completes. >Not every device has an open routine (like network interfaces). They have an attach routine. Maybe the registration could be delayed until the interface is ifconfig'ed. >There should be two steps: the first is registration (allow a driver to >indicate that it will use an IRQ) to detect conflicts, the second is the >actual enabling of the interrupt handler. There aren't enough isa interrupts, and isa interrupts can't be shared (at the same time), so early registration and conflict checking is too restrictive. Of course, interrupts for critical devices need to be reserved early and almost forever. Examples of time-sharing of interrupts: 1) COM1 and COM3 both on IRQ4. Delay the interrupt registration until open() and unregister the interrupt in close(). This would actually be easier to handle using a `controller' device that multiplexes the interrupt internally. 2) Multiple wd disk controllers on IRQ14. Only one can be active at a time. Enable IRQs for the controller that you expect to interrupt next and disable IRQs for the others. Again, this is easier to handle internally. 3) lpt0 on IRQ7 ans something else on IRQ7. Use IRQ7 for lpt0 if nothing else wants it. Otherwise, use IRQ7 for something else and polling for lpt0. This is harder to handle because the something else is probably in another driver and lpt0 needs to be told when its interrupt is preempted. >BSD/OS actually has done this in a manner which is quite nice. There are >no interrupt handlers defined in the config file. Every driver must register >their interrupt handler in their attach routine. >You can almost achieve the same thing in FreeBSD by setting id->id_intr in >the attach routine. I made 2.0 work with existing config files and drivers that don't set id->id_intr. This was too successfull - most drivers still don't set id->id_intr. >The problem is that the irq class is not so easily set >in the driver (because of having isa_devtab_xxx in ioconf.c; it would be >nicer if the isa_device structure had a id_class field which would be used >to register the approriate interrupt handler). I thought of this too, but it was pointed out that the id_class isn't really necessary - it can by implicit in a registration scheme. >Right now, my new-improved DECPA driver needs the following config file line >on FreeBSD: > device di0 at isa? net irq ? Does BSD/OS support the old config keywords that allow you to specify too much in the config file? I'm not sure which keywords are FreeBSD brokenness. Supporting both "irq ?" and "irq n" requires more work by drivers - they should probably honor the request for a particular irq. Some FreeBSD drivers (bt742.c...) just blow away the request for a particular irq. >The "net irq ?" is required so it's put into the right device table. Under >BSD/OS, all I need is: > di0 at isa? Don't you have to register the device as being in class `net'? This would take more than the 4 characters required for `net ' in the config file. Putting it in the config file may be the best approach if the driver is changed more often than the config file (which is far from the case when there are many more installations than driver authors). >The IRQ services in FreeBSD are definitely inferior to BSD/OS. BSD/OS >supports shared interrupts for all interrupts (not specific to PCI as >in FreeBSD). Some routines that FreeBSD could use are: > isa_irqmask_t isa_irqalloc(isa_irqmask_t irqs); > isa_irqmask_t isa_forceintr(int (*forceintr)(void *arg)); Yes. Does BSD/OS use prototypes without __P(()) for everything and good typedefs for everything? The isa irq mask type is magic but since the machine type is fixed, an explicit type can work OK. OTOH, irqalloc() is a good interface for all machines with irq masks and putting isa in its name defeats reuse (does it apply to eisa? :-). forceintr() is a good interface for all machines except its arg should be typedef'ed. forceintr() is difficult to implement. It can't be implemented for lpt's (you need a benign printer attached and turned on). >Modifying my drivers to run under both FreeBSD and BSD/OS has been very >enlightening. Some things BSD/OS does better while FreeBSD does better at >other things. Isn't it easier to have all configuration stuff outside the driver as in FreeBSD? At least if the external configuration is adequate. Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199504250650.QAA19925>