2021-02-17 17:07:17

by Atul Gopinathan

[permalink] [raw]
Subject: [PATCH] staging: comedi: cast to (unsigned int *)

Resolve the following warning generated by sparse:

drivers/staging//comedi/comedi_fops.c:2956:23: warning: incorrect type in assignment (different address spaces)
drivers/staging//comedi/comedi_fops.c:2956:23: expected unsigned int *chanlist
drivers/staging//comedi/comedi_fops.c:2956:23: got void [noderef] <asn:1> *

compat_ptr() has a return type of "void __user *"
as defined in "include/linux/compat.h"

cmd->chanlist is of type "unsigned int *" as defined
in drivers/staging/comedi/comedi.h" in struct
comedi_cmd.

Signed-off-by: Atul Gopinathan <[email protected]>
---
drivers/staging/comedi/comedi_fops.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/staging/comedi/comedi_fops.c b/drivers/staging/comedi/comedi_fops.c
index e85a99b68f31..fc4ec38012b4 100644
--- a/drivers/staging/comedi/comedi_fops.c
+++ b/drivers/staging/comedi/comedi_fops.c
@@ -2953,7 +2953,7 @@ static int get_compat_cmd(struct comedi_cmd *cmd,
cmd->scan_end_arg = v32.scan_end_arg;
cmd->stop_src = v32.stop_src;
cmd->stop_arg = v32.stop_arg;
- cmd->chanlist = compat_ptr(v32.chanlist);
+ cmd->chanlist = (unsigned int __force *)compat_ptr(v32.chanlist);
cmd->chanlist_len = v32.chanlist_len;
cmd->data = compat_ptr(v32.data);
cmd->data_len = v32.data_len;
--
2.27.0


2021-02-17 22:20:08

by Greg KH

[permalink] [raw]
Subject: Re: [PATCH] staging: comedi: cast to (unsigned int *)

On Wed, Feb 17, 2021 at 10:29:08PM +0530, Atul Gopinathan wrote:
> Resolve the following warning generated by sparse:
>
> drivers/staging//comedi/comedi_fops.c:2956:23: warning: incorrect type in assignment (different address spaces)
> drivers/staging//comedi/comedi_fops.c:2956:23: expected unsigned int *chanlist
> drivers/staging//comedi/comedi_fops.c:2956:23: got void [noderef] <asn:1> *
>
> compat_ptr() has a return type of "void __user *"
> as defined in "include/linux/compat.h"
>
> cmd->chanlist is of type "unsigned int *" as defined
> in drivers/staging/comedi/comedi.h" in struct
> comedi_cmd.
>
> Signed-off-by: Atul Gopinathan <[email protected]>
> ---
> drivers/staging/comedi/comedi_fops.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/staging/comedi/comedi_fops.c b/drivers/staging/comedi/comedi_fops.c
> index e85a99b68f31..fc4ec38012b4 100644
> --- a/drivers/staging/comedi/comedi_fops.c
> +++ b/drivers/staging/comedi/comedi_fops.c
> @@ -2953,7 +2953,7 @@ static int get_compat_cmd(struct comedi_cmd *cmd,
> cmd->scan_end_arg = v32.scan_end_arg;
> cmd->stop_src = v32.stop_src;
> cmd->stop_arg = v32.stop_arg;
> - cmd->chanlist = compat_ptr(v32.chanlist);
> + cmd->chanlist = (unsigned int __force *)compat_ptr(v32.chanlist);

__force? That feels wrong, something is odd if that is ever needed.

Are you _sure_ this is correct?

greg k-h

2021-02-17 22:24:43

by Atul Gopinathan

[permalink] [raw]
Subject: Re: [PATCH] staging: comedi: cast to (unsigned int *)

On Wed, Feb 17, 2021 at 06:35:15PM +0100, Greg KH wrote:
> On Wed, Feb 17, 2021 at 10:29:08PM +0530, Atul Gopinathan wrote:
> > Resolve the following warning generated by sparse:
> >
> > drivers/staging//comedi/comedi_fops.c:2956:23: warning: incorrect type in assignment (different address spaces)
> > drivers/staging//comedi/comedi_fops.c:2956:23: expected unsigned int *chanlist
> > drivers/staging//comedi/comedi_fops.c:2956:23: got void [noderef] <asn:1> *
> >
> > compat_ptr() has a return type of "void __user *"
> > as defined in "include/linux/compat.h"
> >
> > cmd->chanlist is of type "unsigned int *" as defined
> > in drivers/staging/comedi/comedi.h" in struct
> > comedi_cmd.
> >
> > Signed-off-by: Atul Gopinathan <[email protected]>
> > ---
> > drivers/staging/comedi/comedi_fops.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/staging/comedi/comedi_fops.c b/drivers/staging/comedi/comedi_fops.c
> > index e85a99b68f31..fc4ec38012b4 100644
> > --- a/drivers/staging/comedi/comedi_fops.c
> > +++ b/drivers/staging/comedi/comedi_fops.c
> > @@ -2953,7 +2953,7 @@ static int get_compat_cmd(struct comedi_cmd *cmd,
> > cmd->scan_end_arg = v32.scan_end_arg;
> > cmd->stop_src = v32.stop_src;
> > cmd->stop_arg = v32.stop_arg;
> > - cmd->chanlist = compat_ptr(v32.chanlist);
> > + cmd->chanlist = (unsigned int __force *)compat_ptr(v32.chanlist);
>
> __force? That feels wrong, something is odd if that is ever needed.
>
> Are you _sure_ this is correct?

The same file has instances of "(usigned int __force *)" cast being
used on the same "cmd->chanlist". For reference:

At line 1797 of comedi_fops.c:
1796 /* restore chanlist pointer before copying back */
1797 cmd->chanlist = (unsigned int __force *)user_chanlist;
1798 cmd->data = NULL;

At line 1880:
1879 /* restore chanlist pointer before copying back */
1880 cmd->chanlist = (unsigned int __force *)user_chanlist;
1881 *copy = true;

Here "user_chanlist" is of type "unsigned int __user *".


Or perhaps, I shouldn't be relying on them?

Thanks!
Atul

2021-02-17 22:29:55

by Greg KH

[permalink] [raw]
Subject: Re: [PATCH] staging: comedi: cast to (unsigned int *)

On Wed, Feb 17, 2021 at 11:40:00PM +0530, Atul Gopinathan wrote:
> On Wed, Feb 17, 2021 at 06:35:15PM +0100, Greg KH wrote:
> > On Wed, Feb 17, 2021 at 10:29:08PM +0530, Atul Gopinathan wrote:
> > > Resolve the following warning generated by sparse:
> > >
> > > drivers/staging//comedi/comedi_fops.c:2956:23: warning: incorrect type in assignment (different address spaces)
> > > drivers/staging//comedi/comedi_fops.c:2956:23: expected unsigned int *chanlist
> > > drivers/staging//comedi/comedi_fops.c:2956:23: got void [noderef] <asn:1> *
> > >
> > > compat_ptr() has a return type of "void __user *"
> > > as defined in "include/linux/compat.h"
> > >
> > > cmd->chanlist is of type "unsigned int *" as defined
> > > in drivers/staging/comedi/comedi.h" in struct
> > > comedi_cmd.
> > >
> > > Signed-off-by: Atul Gopinathan <[email protected]>
> > > ---
> > > drivers/staging/comedi/comedi_fops.c | 2 +-
> > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/staging/comedi/comedi_fops.c b/drivers/staging/comedi/comedi_fops.c
> > > index e85a99b68f31..fc4ec38012b4 100644
> > > --- a/drivers/staging/comedi/comedi_fops.c
> > > +++ b/drivers/staging/comedi/comedi_fops.c
> > > @@ -2953,7 +2953,7 @@ static int get_compat_cmd(struct comedi_cmd *cmd,
> > > cmd->scan_end_arg = v32.scan_end_arg;
> > > cmd->stop_src = v32.stop_src;
> > > cmd->stop_arg = v32.stop_arg;
> > > - cmd->chanlist = compat_ptr(v32.chanlist);
> > > + cmd->chanlist = (unsigned int __force *)compat_ptr(v32.chanlist);
> >
> > __force? That feels wrong, something is odd if that is ever needed.
> >
> > Are you _sure_ this is correct?
>
> The same file has instances of "(usigned int __force *)" cast being
> used on the same "cmd->chanlist". For reference:
>
> At line 1797 of comedi_fops.c:
> 1796 /* restore chanlist pointer before copying back */
> 1797 cmd->chanlist = (unsigned int __force *)user_chanlist;
> 1798 cmd->data = NULL;
>
> At line 1880:
> 1879 /* restore chanlist pointer before copying back */
> 1880 cmd->chanlist = (unsigned int __force *)user_chanlist;
> 1881 *copy = true;
>
> Here "user_chanlist" is of type "unsigned int __user *".
>
>
> Or perhaps, I shouldn't be relying on them?

I don't know, it still feels wrong.

Ian, any thoughts?

thanks,

greg k-h

2021-02-18 15:16:08

by Ian Abbott

[permalink] [raw]
Subject: Re: [PATCH] staging: comedi: cast to (unsigned int *)

On 17/02/2021 18:26, Greg KH wrote:
> On Wed, Feb 17, 2021 at 11:40:00PM +0530, Atul Gopinathan wrote:
>> On Wed, Feb 17, 2021 at 06:35:15PM +0100, Greg KH wrote:
>>> On Wed, Feb 17, 2021 at 10:29:08PM +0530, Atul Gopinathan wrote:
>>>> Resolve the following warning generated by sparse:
>>>>
>>>> drivers/staging//comedi/comedi_fops.c:2956:23: warning: incorrect type in assignment (different address spaces)
>>>> drivers/staging//comedi/comedi_fops.c:2956:23: expected unsigned int *chanlist
>>>> drivers/staging//comedi/comedi_fops.c:2956:23: got void [noderef] <asn:1> *
>>>>
>>>> compat_ptr() has a return type of "void __user *"
>>>> as defined in "include/linux/compat.h"
>>>>
>>>> cmd->chanlist is of type "unsigned int *" as defined
>>>> in drivers/staging/comedi/comedi.h" in struct
>>>> comedi_cmd.
>>>>
>>>> Signed-off-by: Atul Gopinathan <[email protected]>
>>>> ---
>>>> drivers/staging/comedi/comedi_fops.c | 2 +-
>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/staging/comedi/comedi_fops.c b/drivers/staging/comedi/comedi_fops.c
>>>> index e85a99b68f31..fc4ec38012b4 100644
>>>> --- a/drivers/staging/comedi/comedi_fops.c
>>>> +++ b/drivers/staging/comedi/comedi_fops.c
>>>> @@ -2953,7 +2953,7 @@ static int get_compat_cmd(struct comedi_cmd *cmd,
>>>> cmd->scan_end_arg = v32.scan_end_arg;
>>>> cmd->stop_src = v32.stop_src;
>>>> cmd->stop_arg = v32.stop_arg;
>>>> - cmd->chanlist = compat_ptr(v32.chanlist);
>>>> + cmd->chanlist = (unsigned int __force *)compat_ptr(v32.chanlist);
>>>
>>> __force? That feels wrong, something is odd if that is ever needed.
>>>
>>> Are you _sure_ this is correct?
>>
>> The same file has instances of "(usigned int __force *)" cast being
>> used on the same "cmd->chanlist". For reference:
>>
>> At line 1797 of comedi_fops.c:
>> 1796 /* restore chanlist pointer before copying back */
>> 1797 cmd->chanlist = (unsigned int __force *)user_chanlist;
>> 1798 cmd->data = NULL;
>>
>> At line 1880:
>> 1879 /* restore chanlist pointer before copying back */
>> 1880 cmd->chanlist = (unsigned int __force *)user_chanlist;
>> 1881 *copy = true;
>>
>> Here "user_chanlist" is of type "unsigned int __user *".
>>
>>
>> Or perhaps, I shouldn't be relying on them?
>
> I don't know, it still feels wrong.
>
> Ian, any thoughts?

It's kind of moot anyway because the patch is outdated. But the reason
for the ___force is that the same `struct comedi_cmd` is used in both
user and kernel contexts. In user contexts, the `chanlist` member
points to user memory and in kernel contexts it points to kernel memory
(copied from userspace).

The sparse tagging of this member has flip-flopped a bit over the years:

* commit 92d0127c9d24 ("Staging: comedi: __user markup on
comedi_fops.c") (May 2010) tagged it as `__user`.

* commit 9be56c643263 ("staging: comedi: comedi.h: remove __user tag
from chanlist") (Sep 2012) removed the `__user` tag.

It is mostly used in a kernel context, for example all the low-level
drivers with `do_cmd` and `do_cmdtest` handlers use it in kernel context.

The alternative would be to have a separate kernel version of this
struct, but it would be mostly identical to the user version apart from
the sparse tagging of this member and perhaps the removal of the unused
`data` and `data_len` members (which need to be kept in the user version
of the struct for compatibility reasons).

--
-=( Ian Abbott <[email protected]> || MEV Ltd. is a company )=-
-=( registered in England & Wales. Regd. number: 02862268. )=-
-=( Regd. addr.: S11 & 12 Building 67, Europa Business Park, )=-
-=( Bird Hall Lane, STOCKPORT, SK3 0XA, UK. || http://www.mev.co.uk )=-

2021-02-19 09:09:16

by David Laight

[permalink] [raw]
Subject: RE: [PATCH] staging: comedi: cast to (unsigned int *)

> It's kind of moot anyway because the patch is outdated. But the reason
> for the ___force is that the same `struct comedi_cmd` is used in both
> user and kernel contexts. In user contexts, the `chanlist` member
> points to user memory and in kernel contexts it points to kernel memory
> (copied from userspace).

Can't you use a union of the user and kernel pointers?
(Possibly even anonymous?)
Although, ideally, keeping them in separate fields is better.
8 bytes for a pointer isn't going make a fat lot of difference.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)

2021-02-19 09:29:52

by Dan Carpenter

[permalink] [raw]
Subject: Re: [PATCH] staging: comedi: cast to (unsigned int *)

On Fri, Feb 19, 2021 at 09:03:59AM +0000, David Laight wrote:
> > It's kind of moot anyway because the patch is outdated. But the reason
> > for the ___force is that the same `struct comedi_cmd` is used in both
> > user and kernel contexts. In user contexts, the `chanlist` member
> > points to user memory and in kernel contexts it points to kernel memory
> > (copied from userspace).
>
> Can't you use a union of the user and kernel pointers?
> (Possibly even anonymous?)
> Although, ideally, keeping them in separate fields is better.
> 8 bytes for a pointer isn't going make a fat lot of difference.
>

Creating a union is worse than adding casts. With the casts, at least
you know that you're doing something dangerous. It's good that it looks
scary because it is scary.

Keeping them in separate fields is a good idea, but this is part of the
user space API so it's not possible.

The best we can do is adding some more comments so people know why we
are doing the scary casts.

regards,
dan carpenter

2021-02-19 09:40:04

by David Laight

[permalink] [raw]
Subject: RE: [PATCH] staging: comedi: cast to (unsigned int *)

From: Dan Carpenter
> Sent: 19 February 2021 09:26
>
> On Fri, Feb 19, 2021 at 09:03:59AM +0000, David Laight wrote:
> > > It's kind of moot anyway because the patch is outdated. But the reason
> > > for the ___force is that the same `struct comedi_cmd` is used in both
> > > user and kernel contexts. In user contexts, the `chanlist` member
> > > points to user memory and in kernel contexts it points to kernel memory
> > > (copied from userspace).
> >
> > Can't you use a union of the user and kernel pointers?
> > (Possibly even anonymous?)
> > Although, ideally, keeping them in separate fields is better.
> > 8 bytes for a pointer isn't going make a fat lot of difference.
> >
>
> Creating a union is worse than adding casts. With the casts, at least
> you know that you're doing something dangerous. It's good that it looks
> scary because it is scary.
>
> Keeping them in separate fields is a good idea, but this is part of the
> user space API so it's not possible.
>
> The best we can do is adding some more comments so people know why we
> are doing the scary casts.

Another option is to use a longer structure in the kernel with the kernel
pointer in the 'extension'.
So you could have:
struct kernel_foo {
struct foo;
void *kernel_pointer;
};

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)

2021-02-19 09:44:11

by Ian Abbott

[permalink] [raw]
Subject: Re: [PATCH] staging: comedi: cast to (unsigned int *)

On 19/02/2021 09:03, David Laight wrote:
>> It's kind of moot anyway because the patch is outdated. But the reason
>> for the ___force is that the same `struct comedi_cmd` is used in both
>> user and kernel contexts. In user contexts, the `chanlist` member
>> points to user memory and in kernel contexts it points to kernel memory
>> (copied from userspace).
>
> Can't you use a union of the user and kernel pointers?
> (Possibly even anonymous?)
> Although, ideally, keeping them in separate fields is better.
> 8 bytes for a pointer isn't going make a fat lot of difference.

This is for a UAPI header (eventually), so cannot add a new field. For
an anonymous union, one tagged with __user and one not, the __user tag
would be removed during conversion from UAPI headers to
/usr/include/linux headers, leaving a union of two identically typed
members, which would look a bit odd. The union also kind of hides the
problem.

--
-=( Ian Abbott <[email protected]> || MEV Ltd. is a company )=-
-=( registered in England & Wales. Regd. number: 02862268. )=-
-=( Regd. addr.: S11 & 12 Building 67, Europa Business Park, )=-
-=( Bird Hall Lane, STOCKPORT, SK3 0XA, UK. || http://www.mev.co.uk )=-