Is sysfs the right place to get cache and CPU topology info?

Is sysfs the right place to get cache and CPU topology info?

[Paul Mackerras]

I'm being asked by library developers about what is the best way for
them to get information about the CPU cache structure and the CPU
topology from the kernel.  When I said "x86 puts it in sysfs and we'll
do the same on powerpc" I got a response pointing me at this statement
in Documentation/sysfs-rules.txt (the first paragraph):

"The kernel-exported sysfs exports internal kernel implementation details
 and depends on internal kernel structures and layout. It is agreed upon
 by the kernel developers that the Linux kernel does not provide a stable
 internal API. As sysfs is a direct export of kernel internal
 structures, the sysfs interface cannot provide a stable interface either;
 it may always change along with internal kernel changes."

They read that to mean that sysfs is not a suitable interface for them
to use to get information about the system.  In particular they read
that to mean that if they do code their library to read sysfs, it will
change in the future in such a way as to break their code.

In other words, they see sysfs as being completely useless for them
because they can't depend on it as a stable interface.  Which is
reasonable given the quoted paragraph, but on the other hand, I don't
believe we break userspace interfaces as blithely as that paragraph
suggests.

So which is it?  Can they rely on the CPU cache and topology
information under /sys/devices/system/cpu/cpu*, and rely on having
that information there essentially forever?  Or are they correct in
saying sysfs is useless and we need to find some other way to expose
the cache and topology information?

Thanks,
Paul.

Re: Is sysfs the right place to get cache and CPU topology info?

[Andrew Morton]

On Wed, 2 Jul 2008 16:27:25 +1000 Paul Mackerras <paulus@samba.org> wrote:

> I'm being asked by library developers about what is the best way for
> them to get information about the CPU cache structure and the CPU
> topology from the kernel.  When I said "x86 puts it in sysfs and we'll
> do the same on powerpc" I got a response pointing me at this statement
> in Documentation/sysfs-rules.txt (the first paragraph):
> 
> "The kernel-exported sysfs exports internal kernel implementation details
>  and depends on internal kernel structures and layout. It is agreed upon
>  by the kernel developers that the Linux kernel does not provide a stable
>  internal API. As sysfs is a direct export of kernel internal
>  structures, the sysfs interface cannot provide a stable interface either;
>  it may always change along with internal kernel changes."

Those are dopey weasel words and they should be removed.

If we put stuff in sysfs then people WILL use it and we WILL need to
support it for ever.  Pointing at some document and saying "call my
lawyer" just won't cut it.

sysfs is part of the kernel ABI.  We should design our interfaces there
as carefully as we design any others.

> They read that to mean that sysfs is not a suitable interface for them
> to use to get information about the system.  In particular they read
> that to mean that if they do code their library to read sysfs, it will
> change in the future in such a way as to break their code.
> 
> In other words, they see sysfs as being completely useless for them
> because they can't depend on it as a stable interface.  Which is
> reasonable given the quoted paragraph, but on the other hand, I don't
> believe we break userspace interfaces as blithely as that paragraph
> suggests.

Well it's up to them - they own the files.  If they later change them
and break their own interfaces (and presumably their own applications),
well, perhaps they have chosen an inappropriate career?

> So which is it?  Can they rely on the CPU cache and topology
> information under /sys/devices/system/cpu/cpu*, and rely on having
> that information there essentially forever?  Or are they correct in
> saying sysfs is useless and we need to find some other way to expose
> the cache and topology information?

Use sysfs.  Choose a representation which is maitainable in a
backward-compatible fashion for all time.  Maintain it.

Re: Is sysfs the right place to get cache and CPU topology info?

[Paul Mackerras]

Andrew Morton writes:

> Those are dopey weasel words and they should be removed.

Thanks.  That is my opinion too.

> If we put stuff in sysfs then people WILL use it and we WILL need to
> support it for ever.  Pointing at some document and saying "call my
> lawyer" just won't cut it.
> 
> sysfs is part of the kernel ABI.  We should design our interfaces there
> as carefully as we design any others.
> 
> > They read that to mean that sysfs is not a suitable interface for them
> > to use to get information about the system.  In particular they read
> > that to mean that if they do code their library to read sysfs, it will
> > change in the future in such a way as to break their code.
> > 
> > In other words, they see sysfs as being completely useless for them
> > because they can't depend on it as a stable interface.  Which is
> > reasonable given the quoted paragraph, but on the other hand, I don't
> > believe we break userspace interfaces as blithely as that paragraph
> > suggests.
> 
> Well it's up to them - they own the files.  If they later change them
> and break their own interfaces (and presumably their own applications),
> well, perhaps they have chosen an inappropriate career?

We have too many "they"s, perhaps.  I meant that these developers (of
an HPC library that wants to know about cpu caches and topology) see
sysfs as being completely useless as a source of information because
they expect random kernel developers to keep changing it in
incompatible ways.  So "they" (library developers) don't own the files
- they're not kernel developers at all.

> > So which is it?  Can they rely on the CPU cache and topology
> > information under /sys/devices/system/cpu/cpu*, and rely on having
> > that information there essentially forever?  Or are they correct in
> > saying sysfs is useless and we need to find some other way to expose
> > the cache and topology information?
> 
> Use sysfs.  Choose a representation which is maitainable in a
> backward-compatible fashion for all time.  Maintain it.

Thanks. :)

Paul.

Re: Is sysfs the right place to get cache and CPU topology info?

[Andrew Morton]

On Wed, 2 Jul 2008 19:45:43 +1000 Paul Mackerras <paulus@samba.org> wrote:

> > Well it's up to them - they own the files.  If they later change them
> > and break their own interfaces (and presumably their own applications),
> > well, perhaps they have chosen an inappropriate career?
> 
> We have too many "they"s, perhaps.  I meant that these developers (of
> an HPC library that wants to know about cpu caches and topology) see
> sysfs as being completely useless as a source of information because
> they expect random kernel developers to keep changing it in
> incompatible ways.  So "they" (library developers) don't own the files
> - they're not kernel developers at all.

Oh.  I thought "they" (or you) were proposing adding some new
topology-exporting files to sysfs.

If they're talking about using the existing ones then sure, those are
cast in stone as far as I'm concerned.

But they do need to be a _bit_ defensive.  If they see a file which has
multiple name:value fields (shouldn't happen) then don't fail if new
tuples turn up later on.  Don't expect them to always be in the same
order.  Don't fail if new files later turn up in a sysfs directory.  If
a file has (a stupid) format like /proc/self/stat then be prepared for
new columns to appear later on, etc.

But if basic and obvious steps like that are taken in the library, and
later kernel changes cause that library to break, we get to fix the
kernel to unbreak their library.

Re: Is sysfs the right place to get cache and CPU topology info?

[Paul Mackerras]

Andrew Morton writes:

> Oh.  I thought "they" (or you) were proposing adding some new
> topology-exporting files to sysfs.

I was proposing adding them on powerpc using the same format and
location as x86 already uses.  That way the HPC library can use the
one parsing routine on both x86 and powerpc.

> If they're talking about using the existing ones then sure, those are
> cast in stone as far as I'm concerned.
> 
> But they do need to be a _bit_ defensive.  If they see a file which has
> multiple name:value fields (shouldn't happen) then don't fail if new
> tuples turn up later on.  Don't expect them to always be in the same
> order.  Don't fail if new files later turn up in a sysfs directory.  If
> a file has (a stupid) format like /proc/self/stat then be prepared for
> new columns to appear later on, etc.
> 
> But if basic and obvious steps like that are taken in the library, and
> later kernel changes cause that library to break, we get to fix the
> kernel to unbreak their library.

I assume they can rely on finding the stuff they need under
/sys/devices/system/cpu.  Or do they need to traverse the whole of
/sys, and if so, how would they know which directories they should be
looking in?

Paul.

Re: Is sysfs the right place to get cache and CPU topology info?

[Andrew Morton]

On Wed, 2 Jul 2008 21:46:47 +1000 Paul Mackerras <paulus@samba.org> wrote:

> > If they're talking about using the existing ones then sure, those are
> > cast in stone as far as I'm concerned.
> > 
> > But they do need to be a _bit_ defensive.  If they see a file which has
> > multiple name:value fields (shouldn't happen) then don't fail if new
> > tuples turn up later on.  Don't expect them to always be in the same
> > order.  Don't fail if new files later turn up in a sysfs directory.  If
> > a file has (a stupid) format like /proc/self/stat then be prepared for
> > new columns to appear later on, etc.
> > 
> > But if basic and obvious steps like that are taken in the library, and
> > later kernel changes cause that library to break, we get to fix the
> > kernel to unbreak their library.
> 
> I assume they can rely on finding the stuff they need under
> /sys/devices/system/cpu.  Or do they need to traverse the whole of
> /sys, and if so, how would they know which directories they should be
> looking in?

/sys/devices/system/cpu sounds good to me.  Everyone's mounting it at
/sys.

Re: Is sysfs the right place to get cache and CPU topology info?

[Andi Kleen]

Andrew Morton <akpm@linux-foundation.org> writes:
>
> sysfs is part of the kernel ABI.  We should design our interfaces there
> as carefully as we design any others.

The basic problem is that sysfs exports an internal kernel object model
and these tend to change. To really make it stable would require 
splitting it into internal and presented interface.  I would be all
for it, but it doesn't seem realistic to me currently. If we cannot
even get basic interfaces like the syscall capability stable how would
you expect to stabilize the complete kobjects? 

And the specific problem with the x86 cache sysfs interface is that it's so
complicated that no human can really read it directly. This means to
actually use it you need some kind of frontend (i have a cheesy
lscache script for this). I expect that eventually we'll have a standard
tool for this. Right now most people still rely on /proc/cpuinfo
output (which is actually human readable!), but it only shows simple
cache topologies (L2 only) and with L3 and more complicated ones
being more wide spread that doesn't cut it anymore. So I expect
eventually utils-linux will grow a standard lscache program for this.
And I expect people will eventually just use the output frontend instead of 
sysfs.

-Andi

Re: Is sysfs the right place to get cache and CPU topology info?

[Nathan Lynch]

Andi Kleen wrote:
> Andrew Morton <akpm@linux-foundation.org> writes:
> >
> > sysfs is part of the kernel ABI.  We should design our interfaces there
> > as carefully as we design any others.
> 
> The basic problem is that sysfs exports an internal kernel object model
> and these tend to change. To really make it stable would require 
> splitting it into internal and presented interface.

True, but... /sys/devices/system/cpu has been there since around 2.6.5
iirc.  A google code search for that path shows plenty of programs
(including hal) that hard-code it.  Exposed object model or not,
changing that path would break lots of software.


> I would be all
> for it, but it doesn't seem realistic to me currently. If we cannot
> even get basic interfaces like the syscall capability stable how would
> you expect to stabilize the complete kobjects? 
> 
> And the specific problem with the x86 cache sysfs interface is that it's so
> complicated that no human can really read it directly. This means to
> actually use it you need some kind of frontend (i have a cheesy
> lscache script for this).

Human readability is nice, but a more important issue IMO is whether
the cache interface can be considered stable enough for programs to
rely on it.  I notice there's no entry for it in Documentation/ABI.

Re: Is sysfs the right place to get cache and CPU topology info?

[Andi Kleen]

Nathan Lynch wrote:
> Andi Kleen wrote:
>> Andrew Morton <akpm@linux-foundation.org> writes:
>>> sysfs is part of the kernel ABI.  We should design our interfaces there
>>> as carefully as we design any others.
>> The basic problem is that sysfs exports an internal kernel object model
>> and these tend to change. To really make it stable would require 
>> splitting it into internal and presented interface.
> 
> True, but... /sys/devices/system/cpu has been there since around 2.6.5
> iirc.  A google code search for that path shows plenty of programs
> (including hal) that hard-code it.  Exposed object model or not,
> changing that path would break lots of software.

Yes it would.

But Greg is making noises of getting rid of sysdevs and it wouldn't
surprise me if that ended up being user visible since most object
model changes end up being visible.

> 
> 
>> I would be all
>> for it, but it doesn't seem realistic to me currently. If we cannot
>> even get basic interfaces like the syscall capability stable how would
>> you expect to stabilize the complete kobjects? 
>>
>> And the specific problem with the x86 cache sysfs interface is that it's so
>> complicated that no human can really read it directly. This means to
>> actually use it you need some kind of frontend (i have a cheesy
>> lscache script for this).
> 
> Human readability is nice, 

When it's not human readable, then it's usually also not Joe Normal Programmer
usable.

> but a more important issue IMO is whether
> the cache interface can be considered stable enough for programs to
> rely on it.  I notice there's no entry for it in Documentation/ABI.

For x86 it follows closely the Intel CPUID  architectural specification, which
can be considered stable (at least for Intel, other vendors do not
necessarily implement it. But for AMD it is faked at least)

-Andi

Re: Is sysfs the right place to get cache and CPU topology info?

[Greg KH]

On Wed, Jul 02, 2008 at 05:14:02PM +0200, Andi Kleen wrote:
> Nathan Lynch wrote:
> > Andi Kleen wrote:
> >> Andrew Morton <akpm@linux-foundation.org> writes:
> >>> sysfs is part of the kernel ABI.  We should design our interfaces there
> >>> as carefully as we design any others.
> >> The basic problem is that sysfs exports an internal kernel object model
> >> and these tend to change. To really make it stable would require 
> >> splitting it into internal and presented interface.
> > 
> > True, but... /sys/devices/system/cpu has been there since around 2.6.5
> > iirc.  A google code search for that path shows plenty of programs
> > (including hal) that hard-code it.  Exposed object model or not,
> > changing that path would break lots of software.
> 
> Yes it would.
> 
> But Greg is making noises of getting rid of sysdevs and it wouldn't
> surprise me if that ended up being user visible since most object
> model changes end up being visible.

I hope to make sysdevs go away in such a manner that the sysfs tree does
not change at all.  That's my goal, but we still have a long ways to go
before we can even consider attempting to do this, so don't worry about
putting things in this location if you feel it is the best fit.

thanks,

greg k-h

Removing sysdevs? (was: Re: Is sysfs the right place to get cache and CPU topology info?)

[Rafael J. Wysocki]

On Wednesday, 2 of July 2008, Greg KH wrote:
> On Wed, Jul 02, 2008 at 05:14:02PM +0200, Andi Kleen wrote:
> > Nathan Lynch wrote:
> > > Andi Kleen wrote:
> > >> Andrew Morton <akpm@linux-foundation.org> writes:
> > >>> sysfs is part of the kernel ABI.  We should design our interfaces there
> > >>> as carefully as we design any others.
> > >> The basic problem is that sysfs exports an internal kernel object model
> > >> and these tend to change. To really make it stable would require 
> > >> splitting it into internal and presented interface.
> > > 
> > > True, but... /sys/devices/system/cpu has been there since around 2.6.5
> > > iirc.  A google code search for that path shows plenty of programs
> > > (including hal) that hard-code it.  Exposed object model or not,
> > > changing that path would break lots of software.
> > 
> > Yes it would.
> > 
> > But Greg is making noises of getting rid of sysdevs and it wouldn't
> > surprise me if that ended up being user visible since most object
> > model changes end up being visible.
> 
> I hope to make sysdevs go away in such a manner that the sysfs tree does
> not change at all.  That's my goal, but we still have a long ways to go
> before we can even consider attempting to do this, so don't worry about
> putting things in this location if you feel it is the best fit.

Speaking of which, I'm very interested in the removing of sysdevs, since they
don't fit into the new suspend/hibernation framework I'm working on.  Can you
please tell me what the plan is?

Thanks,
Rafael

Re: Removing sysdevs? (was: Re: Is sysfs the right place to get cache and CPU topology info?)

[Greg KH]

On Wed, Jul 02, 2008 at 11:41:44PM +0200, Rafael J. Wysocki wrote:
> On Wednesday, 2 of July 2008, Greg KH wrote:
> > On Wed, Jul 02, 2008 at 05:14:02PM +0200, Andi Kleen wrote:
> > > Nathan Lynch wrote:
> > > > Andi Kleen wrote:
> > > >> Andrew Morton <akpm@linux-foundation.org> writes:
> > > >>> sysfs is part of the kernel ABI.  We should design our interfaces there
> > > >>> as carefully as we design any others.
> > > >> The basic problem is that sysfs exports an internal kernel object model
> > > >> and these tend to change. To really make it stable would require 
> > > >> splitting it into internal and presented interface.
> > > > 
> > > > True, but... /sys/devices/system/cpu has been there since around 2.6.5
> > > > iirc.  A google code search for that path shows plenty of programs
> > > > (including hal) that hard-code it.  Exposed object model or not,
> > > > changing that path would break lots of software.
> > > 
> > > Yes it would.
> > > 
> > > But Greg is making noises of getting rid of sysdevs and it wouldn't
> > > surprise me if that ended up being user visible since most object
> > > model changes end up being visible.
> > 
> > I hope to make sysdevs go away in such a manner that the sysfs tree does
> > not change at all.  That's my goal, but we still have a long ways to go
> > before we can even consider attempting to do this, so don't worry about
> > putting things in this location if you feel it is the best fit.
> 
> Speaking of which, I'm very interested in the removing of sysdevs, since they
> don't fit into the new suspend/hibernation framework I'm working on.  Can you
> please tell me what the plan is?

The plan is:
	- remaining driver core cleanups to allow for multiple drivers
	  to be bound to individual devices
	- add multiple binding support to the core
	- migrate existing sysdevs to struct device, now that multiple
	  binding is allowed
	- delete sysdev structure
	- profit!

It's that first step that is taking a while, the last big changes will
be going into 2.6.27 to help accomplish this, after that merge happens
for 2.6.27-rc1 I'll be working on the remaining steps.

thanks,

greg k-h

Re: Removing sysdevs?

[Andi Kleen]

> The plan is:
> 	- remaining driver core cleanups to allow for multiple drivers
> 	  to be bound to individual devices
> 	- add multiple binding support to the core
> 	- migrate existing sysdevs to struct device, now that multiple
> 	  binding is allowed
> 	- delete sysdev structure
> 	- profit!

I hope the sysdev semantics of running with interrupts etc. will be kept?

> 
> It's that first step that is taking a while, the last big changes will
> be going into 2.6.27 to help accomplish this, after that merge happens
> for 2.6.27-rc1 I'll be working on the remaining steps.

So what's up with the sysdev attribute passing changes i posted? Are you
going to put them in or will the conversion happen so soon that they're
not worth it?

I eventually need the passed attribute for the dynamic bank patches.

-Andi

Re: Removing sysdevs?

[Greg KH]

On Wed, Jul 02, 2008 at 11:57:00PM +0200, Andi Kleen wrote:
> 
> > The plan is:
> > 	- remaining driver core cleanups to allow for multiple drivers
> > 	  to be bound to individual devices
> > 	- add multiple binding support to the core
> > 	- migrate existing sysdevs to struct device, now that multiple
> > 	  binding is allowed
> > 	- delete sysdev structure
> > 	- profit!
> 
> I hope the sysdev semantics of running with interrupts etc. will be kept?

Yes, all of the current semantics should be kept.

> > It's that first step that is taking a while, the last big changes will
> > be going into 2.6.27 to help accomplish this, after that merge happens
> > for 2.6.27-rc1 I'll be working on the remaining steps.
> 
> So what's up with the sysdev attribute passing changes i posted? Are you
> going to put them in or will the conversion happen so soon that they're
> not worth it?
> 
> I eventually need the passed attribute for the dynamic bank patches.

Sorry, was side tracked by other work today.

The patches look fine, do you want me to take them through my tree, or
do you want to have them go through somewhere else as you have some
dependancies on them?

If somewhere else, that's fine with me, feel free to add:
	Acked-by: Greg Kroah-Hartman <gregkh@suse.de>
to them.

thanks,

greg k-h

Re: Removing sysdevs?

[Andi Kleen]

> The patches look fine, do you want me to take them through my tree, or
> do you want to have them go through somewhere else as you have some
> dependancies on them?

Through your tree is fine. Thanks.

Perhaps drop the x86 dynamic bank patch, it would actually conflict
with one patch currently in the x86 tree and doesn't really
belong into your tree.

> 
> If somewhere else, that's fine with me, feel free to add:
> 	Acked-by: Greg Kroah-Hartman <gregkh@suse.de>
> to them.

-Andi

Re: Removing sysdevs?

[Greg KH]

On Thu, Jul 03, 2008 at 11:53:14AM +0200, Andi Kleen wrote:
> 
> > The patches look fine, do you want me to take them through my tree, or
> > do you want to have them go through somewhere else as you have some
> > dependancies on them?
> 
> Through your tree is fine. Thanks.
> 
> Perhaps drop the x86 dynamic bank patch, it would actually conflict
> with one patch currently in the x86 tree and doesn't really
> belong into your tree.

Dropped, I only applied the other 3.

thanks,

greg k-h

Re: Removing sysdevs?

[Benjamin Herrenschmidt]

On Wed, 2008-07-02 at 23:57 +0200, Andi Kleen wrote:
> > The plan is:
> > 	- remaining driver core cleanups to allow for multiple drivers
> > 	  to be bound to individual devices
> > 	- add multiple binding support to the core
> > 	- migrate existing sysdevs to struct device, now that multiple
> > 	  binding is allowed
> > 	- delete sysdev structure
> > 	- profit!
> 
> I hope the sysdev semantics of running with interrupts etc. will be kept?

What "running with interrupts" ? :-) Are you talking specifically about
suspend/resume ? In this case, normal devices also provide irq_off
variants of suspend/resume.

In some areas, those semantics are even a problem for sysdevs. For
example, it's been a pain in the neck for cpufreq on powermac not to be
able to schedule in its suspend routine. Among others...

Cheers,
Ben.

Re: Removing sysdevs? (was: Re: Is sysfs the right place to get cache and CPU topology info?)

[Rafael J. Wysocki]

On Wednesday, 2 of July 2008, Greg KH wrote:
> On Wed, Jul 02, 2008 at 11:41:44PM +0200, Rafael J. Wysocki wrote:
> > On Wednesday, 2 of July 2008, Greg KH wrote:
> > > On Wed, Jul 02, 2008 at 05:14:02PM +0200, Andi Kleen wrote:
> > > > Nathan Lynch wrote:
> > > > > Andi Kleen wrote:
> > > > >> Andrew Morton <akpm@linux-foundation.org> writes:
> > > > >>> sysfs is part of the kernel ABI.  We should design our interfaces there
> > > > >>> as carefully as we design any others.
> > > > >> The basic problem is that sysfs exports an internal kernel object model
> > > > >> and these tend to change. To really make it stable would require 
> > > > >> splitting it into internal and presented interface.
> > > > > 
> > > > > True, but... /sys/devices/system/cpu has been there since around 2.6.5
> > > > > iirc.  A google code search for that path shows plenty of programs
> > > > > (including hal) that hard-code it.  Exposed object model or not,
> > > > > changing that path would break lots of software.
> > > > 
> > > > Yes it would.
> > > > 
> > > > But Greg is making noises of getting rid of sysdevs and it wouldn't
> > > > surprise me if that ended up being user visible since most object
> > > > model changes end up being visible.
> > > 
> > > I hope to make sysdevs go away in such a manner that the sysfs tree does
> > > not change at all.  That's my goal, but we still have a long ways to go
> > > before we can even consider attempting to do this, so don't worry about
> > > putting things in this location if you feel it is the best fit.
> > 
> > Speaking of which, I'm very interested in the removing of sysdevs, since they
> > don't fit into the new suspend/hibernation framework I'm working on.  Can you
> > please tell me what the plan is?
> 
> The plan is:
> 	- remaining driver core cleanups to allow for multiple drivers
> 	  to be bound to individual devices
> 	- add multiple binding support to the core
> 	- migrate existing sysdevs to struct device, now that multiple
> 	  binding is allowed

Once they've been migrated to struct device, will they reside on specific
'system' bus, or will they be platform devices?

> 	- delete sysdev structure
> 	- profit!
> 
> It's that first step that is taking a while, the last big changes will
> be going into 2.6.27 to help accomplish this, after that merge happens
> for 2.6.27-rc1 I'll be working on the remaining steps.

Sounds good, please let me know if you need help.

Thanks,
Rafael

Re: Removing sysdevs? (was: Re: Is sysfs the right place to get cache and CPU topology info?)

[Greg KH]

On Thu, Jul 03, 2008 at 12:08:45AM +0200, Rafael J. Wysocki wrote:
> On Wednesday, 2 of July 2008, Greg KH wrote:
> > On Wed, Jul 02, 2008 at 11:41:44PM +0200, Rafael J. Wysocki wrote:
> > > On Wednesday, 2 of July 2008, Greg KH wrote:
> > > > On Wed, Jul 02, 2008 at 05:14:02PM +0200, Andi Kleen wrote:
> > > > > Nathan Lynch wrote:
> > > > > > Andi Kleen wrote:
> > > > > >> Andrew Morton <akpm@linux-foundation.org> writes:
> > > > > >>> sysfs is part of the kernel ABI.  We should design our interfaces there
> > > > > >>> as carefully as we design any others.
> > > > > >> The basic problem is that sysfs exports an internal kernel object model
> > > > > >> and these tend to change. To really make it stable would require 
> > > > > >> splitting it into internal and presented interface.
> > > > > > 
> > > > > > True, but... /sys/devices/system/cpu has been there since around 2.6.5
> > > > > > iirc.  A google code search for that path shows plenty of programs
> > > > > > (including hal) that hard-code it.  Exposed object model or not,
> > > > > > changing that path would break lots of software.
> > > > > 
> > > > > Yes it would.
> > > > > 
> > > > > But Greg is making noises of getting rid of sysdevs and it wouldn't
> > > > > surprise me if that ended up being user visible since most object
> > > > > model changes end up being visible.
> > > > 
> > > > I hope to make sysdevs go away in such a manner that the sysfs tree does
> > > > not change at all.  That's my goal, but we still have a long ways to go
> > > > before we can even consider attempting to do this, so don't worry about
> > > > putting things in this location if you feel it is the best fit.
> > > 
> > > Speaking of which, I'm very interested in the removing of sysdevs, since they
> > > don't fit into the new suspend/hibernation framework I'm working on.  Can you
> > > please tell me what the plan is?
> > 
> > The plan is:
> > 	- remaining driver core cleanups to allow for multiple drivers
> > 	  to be bound to individual devices
> > 	- add multiple binding support to the core
> > 	- migrate existing sysdevs to struct device, now that multiple
> > 	  binding is allowed
> 
> Once they've been migrated to struct device, will they reside on specific
> 'system' bus, or will they be platform devices?

I haven't really thought it through more than the above yet, so I don't
know :)

thanks,

greg k-h

Re: Removing sysdevs? (was: Re: Is sysfs the right place to get cache and CPU topology info?)

[Rafael J. Wysocki]

On Thursday, 3 of July 2008, Greg KH wrote:
> On Thu, Jul 03, 2008 at 12:08:45AM +0200, Rafael J. Wysocki wrote:
> > On Wednesday, 2 of July 2008, Greg KH wrote:
> > > On Wed, Jul 02, 2008 at 11:41:44PM +0200, Rafael J. Wysocki wrote:
> > > > On Wednesday, 2 of July 2008, Greg KH wrote:
> > > > > On Wed, Jul 02, 2008 at 05:14:02PM +0200, Andi Kleen wrote:
> > > > > > Nathan Lynch wrote:
> > > > > > > Andi Kleen wrote:
> > > > > > >> Andrew Morton <akpm@linux-foundation.org> writes:
> > > > > > >>> sysfs is part of the kernel ABI.  We should design our interfaces there
> > > > > > >>> as carefully as we design any others.
> > > > > > >> The basic problem is that sysfs exports an internal kernel object model
> > > > > > >> and these tend to change. To really make it stable would require 
> > > > > > >> splitting it into internal and presented interface.
> > > > > > > 
> > > > > > > True, but... /sys/devices/system/cpu has been there since around 2.6.5
> > > > > > > iirc.  A google code search for that path shows plenty of programs
> > > > > > > (including hal) that hard-code it.  Exposed object model or not,
> > > > > > > changing that path would break lots of software.
> > > > > > 
> > > > > > Yes it would.
> > > > > > 
> > > > > > But Greg is making noises of getting rid of sysdevs and it wouldn't
> > > > > > surprise me if that ended up being user visible since most object
> > > > > > model changes end up being visible.
> > > > > 
> > > > > I hope to make sysdevs go away in such a manner that the sysfs tree does
> > > > > not change at all.  That's my goal, but we still have a long ways to go
> > > > > before we can even consider attempting to do this, so don't worry about
> > > > > putting things in this location if you feel it is the best fit.
> > > > 
> > > > Speaking of which, I'm very interested in the removing of sysdevs, since they
> > > > don't fit into the new suspend/hibernation framework I'm working on.  Can you
> > > > please tell me what the plan is?
> > > 
> > > The plan is:
> > > 	- remaining driver core cleanups to allow for multiple drivers
> > > 	  to be bound to individual devices
> > > 	- add multiple binding support to the core
> > > 	- migrate existing sysdevs to struct device, now that multiple
> > > 	  binding is allowed
> > 
> > Once they've been migrated to struct device, will they reside on specific
> > 'system' bus, or will they be platform devices?
> 
> I haven't really thought it through more than the above yet, so I don't
> know :)

This is quite important, though, because the device objects that sysdevs will
be replaced with should provide suspend/hibernation callbacks to be run with
interrupts disabled, while for some of them it may also be convenient to
provide "normal" suspend/hibernation callbacks to be run with interrupts
enabled.

For this reason their bus type will have to be quite similar to the platform bus
type.

Thanks,
Rafael

[PATCH] sysfs-rules.txt: reword API stability statement

[Nathan Lynch]

The first paragraph of this document implies that user space
developers shouldn't use sysfs at all, but then it goes on to describe
rules that developers should follow when accessing sysfs.  Not only is
this somewhat self-contradictory, it has been shown to discourage
developers from using established sysfs interfaces.

A note of caution is more appropriate than a blanket "sysfs will never
be stable" assertion.

Signed-off-by: Nathan Lynch <ntl@pobox.com>
---

Andrew Morton wrote:
> 
> Those are dopey weasel words and they should be removed.

 Documentation/sysfs-rules.txt |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/Documentation/sysfs-rules.txt b/Documentation/sysfs-rules.txt
index 80ef562..6049a2a 100644
--- a/Documentation/sysfs-rules.txt
+++ b/Documentation/sysfs-rules.txt
@@ -3,9 +3,8 @@ Rules on how to access information in the Linux kernel sysfs
 The kernel-exported sysfs exports internal kernel implementation details
 and depends on internal kernel structures and layout. It is agreed upon
 by the kernel developers that the Linux kernel does not provide a stable
-internal API. As sysfs is a direct export of kernel internal
-structures, the sysfs interface cannot provide a stable interface either;
-it may always change along with internal kernel changes.
+internal API. Therefore, there are aspects of the sysfs interface that
+may not be stable across kernel releases.
 
 To minimize the risk of breaking users of sysfs, which are in most cases
 low-level userspace applications, with a new kernel release, the users
-- 
1.5.5