set_fs removal for s390

set_fs removal for s390

[Christoph Hellwig]

Hi s390 maintainers,

the base work to remove address space overrides using set_fs() has
landed in linux-next through this tree:

https://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git/log/?h=base.set_fs

and in addition to x86 and powerpc converted there, we have arm, arm64
and RISC-V conversion in progress.  That means s390 is the only "tier 1"
architecture still missing.  Due to the special instructions for
accessing the user space address space I don't really feel like doing
the conversion myself, but it would be great if we could get s390 done
for 5.10 as well.

The rought TODO list is:
 (1) actually stop using set_fs in the zrcypt driver.  I could prepare
     a crude untested patch for that if it helps.
 (2) implement the __get_user_nofault and __put_user_nofault helpers to
     safely access kernel memory.  These should behave like
     get_user/put_user under set_fs(KERNEL_DS), but not actually
     manipulate any task state to do so
 (3) unselect CONFIG_SET_FS and remove all the code related to overriding
     the address space(3) unselect CONFIG_SET_FS and remove all the code
     related to overriding the address space limit

Re: set_fs removal for s390

[Heiko Carstens]

On Tue, Sep 08, 2020 at 08:31:33AM +0200, Christoph Hellwig wrote:
> Hi s390 maintainers,
> 
> the base work to remove address space overrides using set_fs() has
> landed in linux-next through this tree:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git/log/?h=base.set_fs
> 
> and in addition to x86 and powerpc converted there, we have arm, arm64
> and RISC-V conversion in progress.  That means s390 is the only "tier 1"
> architecture still missing.  Due to the special instructions for
> accessing the user space address space I don't really feel like doing
> the conversion myself, but it would be great if we could get s390 done
> for 5.10 as well.
> 
> The rought TODO list is:
>  (1) actually stop using set_fs in the zrcypt driver.  I could prepare
>      a crude untested patch for that if it helps.

Hmm.. Harald's initial implementation years ago was without set_fs(),
if I remember correctly, and then he was talked into using set_fs() ;)

Harald, can you take care of this part?

>  (2) implement the __get_user_nofault and __put_user_nofault helpers to
>      safely access kernel memory.  These should behave like
>      get_user/put_user under set_fs(KERNEL_DS), but not actually
>      manipulate any task state to do so
>  (3) unselect CONFIG_SET_FS and remove all the code related to overriding
>      the address space(3) unselect CONFIG_SET_FS and remove all the code
>      related to overriding the address space limit

Ok, I guess everything will have to land in vfs.git then.

Re: set_fs removal for s390

[Christoph Hellwig]

On Tue, Sep 08, 2020 at 02:26:37PM +0200, Heiko Carstens wrote:
> >  (2) implement the __get_user_nofault and __put_user_nofault helpers to
> >      safely access kernel memory.  These should behave like
> >      get_user/put_user under set_fs(KERNEL_DS), but not actually
> >      manipulate any task state to do so
> >  (3) unselect CONFIG_SET_FS and remove all the code related to overriding
> >      the address space(3) unselect CONFIG_SET_FS and remove all the code
> >      related to overriding the address space limit
> 
> Ok, I guess everything will have to land in vfs.git then.

The branch is supposed to be a stable base that you can pull into
your tree.