RE: devfs + PCI serial card = no extra serial ports
Ed Vance (EdV@macrolink.com)
Fri, 14 Mar 2003 11:49:54 -0800
On Fri, Mar 14, 2003 at 11:06 AM, Adam J. Richter wrote:
> On 2003-03-14 at 15:23:50, David Woodhouse wrote:
> >On Mon, 2003-03-10 at 18:25, Bryan Whitehead wrote:
> >> [snip]
> >> > There is nothing in devfs that prevents you from
> >> >registering devfs devices even if they are not yet bound
> >> >to specific hardware (you do not need a sysfs mapping,
> >> >for example). So, you should be able to register
> >> >/dev/tts/0..N at initialization, where N is the
> >> > maximum number of serial devices you want to support.
> >> are you saying there is a way to force devfs to make more
> >> entries in /dev/tts/ without any hardware being attached
> >> to the entries? Then i can use setserial? so on boot I'd
> >> have 4 entries in /dev/tts ?
> >Don't do this. The whole concept of opening a device node
> >for a device which is _absent_, then doing magic ioctls on
> >it to make the driver probe for the hardware, is utterly
> >Fix it properly instead -- disallow opening of a /dev/ttySx
> >node with uart type unknown, and implement a proper way to
> >tell the serial driver 'please look for a device _here_',
> >via sysfs or something.
> I don't know what you mean by "is utterly bogus." You need to
> explain why you think this practice results in a larger kernel
> footprint, lower throughput, longer latency, larger object or source
> code size, or is worse by some other objective external metric when
> compared to a specific alternative (another element you seem to have
> This is a frequent problem that I see in many public technical
> discussions. Using intimidation words like "is utterly bogus" to
> cover for not stating real technical reasons not only wastes peoples
> time in reading an unnecessary email exchange but also can mislead
> people into decision chains that tack toward producing software that
> is bigger, slower, harder to maintain, lacking in functionality or
> worse by most weightings of external benefits that people would
> Getting back to the topic of /dev devices without a constant
> binding to hardware devices, a device having at least a
> context-depenedent binding goes back as far as /dev/tty. Being able
> to define a new serial device with open and ioctl have been in the
> Linux kernel since before devfs ever existed. Non-devfs systems
> generally have many device files in /dev that are not actually bound
> to any existing devcies. For example, a non-devfs system may have
> device files for the first four SCSI disks even on a system that has
> no SCSI disks. In fact, such a device file may become valid later if
> a USB disk is plugged in. User programs seem to be working OK with
> the non-constance of a device file's binding to a specific hardware
> device (i.e., there has not been some big unintended consequence that
> has caused systems lock up when they boot or when it's time to run
> fsck or after 8 hours of use or whatever). So, the costs of this
> approach that I can think of seem pretty small.
> Being able to open a device and then using ioctl's to define
> it (such as is currently the case in serial and loop devices) creates
> a programming interface that is more easily ported to other operating
> systems that don't have /sys and but do have open and ioctl. The
> existing approach is also more capitable with versions of Linux that
> lack /sys.
> Before these ioctl's are run on a serial device to set the
> UART type, interrupt and IO port, the kernel does not necessarily know
> that there is a UART at a particular IO location, what type it is and
> what interrupt it is associated with. So, it sounds like a /sys
> interface would be more kernel code than the current scheme without
> solving any real problem or delivering any real benefit that I see.
> Also, the dynamic configuration of serial ports with setserial
> is code that has been widely used for a long time, so its reliability
> should be pretty good.
> So, it seems to me that the expected reliability,
> compatability, kernel code size and user code size benefits are likely
> to exceed those of defining some new creation process using /sys. If
> you disagree, then please come up with some specific proposal (even if
> just for purposes of example), list these trade-offs and whatever
> others you perceive, and then we can discuss which approach each of
> thinks provides the most benefit on balance.
The largest gotcha that I am aware of for the PCI/setserial command
combination is the inability to automatically "follow" the card when
its address changes due to adding or removing an unrelated PCI device.
Even so, the system is unlikely to crash because the driver checks the
LSR register for impossible values and cuts off access when the UART
is not present.
Setserial was designed in the days of ISA serial cards which did not
have the moving-target issue.
I agree that a better interface for dynamic serial port configuration
is needed. In particular, I would like to see coverage of memory mapped
cards. However (comma) do we want to destabilize 2.4 to make fundamental
changes to the serial driver configuration interfaces? Probably not. If
it's really a 2.5 issue, then we should look at Russell King's redesign
in 2.5 and make constructive suggestions (patches) for 2.5. (_after_
reading the LKML archive threads about it)
Anybody know of any other (worse) technical issues with PCI/setserial
on the 2.4 kernels?
Ed Vance edv (at) macrolink (dot) com
Macrolink, Inc. 1500 N. Kellogg Dr Anaheim, CA 92807
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/