[PATCH 0/2] arm: add a /proc/cpuinfo platform extension

[H Hartley Sweeten]

On Tuesday, March 23, 2010 3:02 PM, Ryan Mallon wrote:
> H Hartley Sweeten wrote:
>> On Tuesday, March 23, 2010 1:30 PM, Ryan Mallon wrote:
>>> H Hartley Sweeten wrote:
>>>> Add an optional platform specific extension to /proc/cpuinfo.
>>>>
>>>> Many platforms have custom cpu information that could be exposed
>>>> to user space using /proc/cpuinfo.
>>>>
>>>> Patch 1/2 adds the necessary core support to allow a platform
>>>> specific callback to dump this information.
>>>>
>>>> Patch 2/2 adds a callback to mach-ep93xx and hooks up all the
>>>> edb93xx platforms.
>>>>
>>>> Signed-off-by: H Hartley Sweeten <hsweeten at visionengravers.com>
>>>>
>>>> _______________________________________________
>>>> linux-arm-kernel mailing list
>>>> linux-arm-kernel at lists.infradead.org
>>>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>>> I think this is unlikely to get merged in its current state. Russell has
>>> mentioned issues with breaking userspace by changing /proc/cpuinfo. 
>> 
>> I don't agree with this point.
>> 
>
> [snip]
>
>> 
>> Even something as trivial as the BogoMIPS is in a different place in
>> the two outputs and is spelled differently (due to caps).
>> 
>> The outputs are completely different.  Other architectures in
>> mainline also have very different outputs.
>> 
>> I can't see any reason why adding additional fields will break
>> user space, as long as an existing heading in the output is not
>> duplicated.  Even that "really" shouldn't break anything since
>> any application parsing this file has to do it sequentially and
>> the new headings are located at the end of the file.
>
> I'm really not sure. There may be some crappy userspace tools out there
> which will break. I don't really mind either way if the info goes in
> /proc/cpuinfo, or some new /proc/archinfo, just as long as it doesn't
> break userspace in some way.

I did a grep of glibc's source (cvs-2.9, the only one currently on my
system) to see what it tries to parse out of /proc/cpuinfo.  These are
the only ones I could find:

./sysdeps/unix/sysv/linux/arm/ioperm.c
	"Hardware\t: %256[^\n]\n"
./sysdeps/unix/sysv/linux/m68k/getsysstats.c
	"CPU:"
./sysdeps/unix/sysv/linux/mips/getsysstats.c
	"cpu model"
./nptl/perf.c
	"cpu MHz"
./nptl/sockperf.c
	"cpu MHz"
./sysdeps/unix/sysv/linux/getsysstats.c
	"processor"
./sysdeps/unix/sysv/linux/alpha/getsysstats.c
	"cpus active : %d"
	"CPUs probed %*d active %d"
	"cpus detected : %d"
	"CPUs probed %d"
./sysdeps/unix/sysv/linux/alpha/ioperm.c
	"system type : %256[^\n]\n"
	"system variation : %256[^\n]\n"
	"cpu model :%256[^\n]\n"
./sysdeps/unix/sysv/linux/i386/get_clockfreq.c
	"cpu MHz"
./sysdeps/unix/sysv/linux/ia64/get_clockfreq.c
	"itc MHz"
./sysdeps/unix/sysv/linux/powerpc/get_clockfreq.c
	"timebase"
./sysdeps/unix/sysv/linux/sparc/getsysstats.c
	"ncpus active : %d"
	"ncpus probed : %d"
./sysdeps/unix/sysv/linux/sparc/sparc64/get_clockfreq.c
	"Cpu0ClkTck"

There may be other userspace tools out there that look at this file
but I would assume they still would parse it using fopen/sscanf/fclose.
Nothing should break as long as the new data is at the end of the file.

>>> The other problem I see is that you have a single callback for registering
>>> the arch specific information. In you ep93xx example, each of the ep93xx
>>> boards must add:
>>>
>>>  .arch_cpuinfo = ep93xx_cpuinfo,
>>>
>>> If one of the boards has some additional information to make available,
>>> it would need to reimplement the entire callback, which gets messy.
>> 
>> Not necessarily.
>> 
>> If a board, such as the ts72xx, wanted to add additional information
>> it just has to register it's private callback then call the ep93xx core
>> supplied callback at the desired point in it's private one.
>> 
>> The ts72xx currently does this exact thing with the .map_io callback.
>> It supplies it's own private one to map the external FPGA.  It first calls
>> the ep93xx core to map the ahb/apb space then it does an iotable_init to
>> map the FPGA.
>
> Okay, fair point. I still don't like having the seq_file callback being
> in machine_desc. It means that all of the board files have to be edited
> to add the callback. It should be something which happens automagically
> in the platform core. Perhaps using a weak function for the callback, or
> a #define check.

The callback is copied from the machine_desc in arch/arm/kernel/setup.c
just like the other architecture-specific pointers.  If an architecture
does not have any additional data to dump they don't have to provide the
callback.  If it's not provided, the main seq_file callback (c_show in
arch/arm/kernel/setup.c) will not call it.  See Patch 1/3 of the series.

Regards,
Hartley