[PATCH] fs: Call kmap_local_page() in copy_string_kernel()

[PATCH] fs: Call kmap_local_page() in copy_string_kernel()

[Fabio M. De Francesco]

The use of kmap_atomic() is being deprecated in favor of kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local, not
globally visible and can take page faults. Furthermore, the mappings can be
acquired from any context (including interrupts).

Therefore, use kmap_local_page() in copy_string_kernel() instead of
kmap_atomic().

Tested with xfstests on a QEMU + KVM 32-bits VM booting a kernel with
HIGHMEM64GB enabled.

Suggested-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Fabio M. De Francesco <fmdefrancesco@gmail.com>
---

I sent a first patch to fs/exec.c for converting kmap() and kmap_atomic()
to kmap_local_page():
https://lore.kernel.org/lkml/20220630163527.9776-1-fmdefrancesco@gmail.com/

Some days ago, Ira Weiny, while he was reviewing that patch, made me notice
that I had overlooked a second kmap_atomic() in the same file (thanks):
https://lore.kernel.org/lkml/YsiQptk19txHrG4c@iweiny-desk3/

I've been asked to send this as an additional change. This is why there will
not be any second version of that previous patch.

 fs/exec.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/exec.c b/fs/exec.c
index 4a2129c0d422..5fa652ca5823 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -639,11 +639,11 @@ int copy_string_kernel(const char *arg, struct linux_binprm *bprm)
 		page = get_arg_page(bprm, pos, 1);
 		if (!page)
 			return -E2BIG;
-		kaddr = kmap_atomic(page);
+		kaddr = kmap_local_page(page);
 		flush_arg_page(bprm, pos & PAGE_MASK, page);
 		memcpy(kaddr + offset_in_page(pos), arg, bytes_to_copy);
 		flush_dcache_page(page);
-		kunmap_atomic(kaddr);
+		kunmap_local(kaddr);
 		put_arg_page(page);
 	}
 
-- 
2.36.1

Re: [PATCH] fs: Call kmap_local_page() in copy_string_kernel()

[Ira Weiny]

On Sun, Jul 10, 2022 at 12:01:36PM +0200, Fabio M. De Francesco wrote:
> The use of kmap_atomic() is being deprecated in favor of kmap_local_page().
> 
> With kmap_local_page(), the mappings are per thread, CPU local, not
> globally visible and can take page faults. Furthermore, the mappings can be
> acquired from any context (including interrupts).
> 
> Therefore, use kmap_local_page() in copy_string_kernel() instead of
> kmap_atomic().
> 
> Tested with xfstests on a QEMU + KVM 32-bits VM booting a kernel with
> HIGHMEM64GB enabled.
> 
> Suggested-by: Ira Weiny <ira.weiny@intel.com>
> Signed-off-by: Fabio M. De Francesco <fmdefrancesco@gmail.com>
> ---
> 
> I sent a first patch to fs/exec.c for converting kmap() and kmap_atomic()
> to kmap_local_page():
> https://lore.kernel.org/lkml/20220630163527.9776-1-fmdefrancesco@gmail.com/
> 
> Some days ago, Ira Weiny, while he was reviewing that patch, made me notice
> that I had overlooked a second kmap_atomic() in the same file (thanks):
> https://lore.kernel.org/lkml/YsiQptk19txHrG4c@iweiny-desk3/
> 
> I've been asked to send this as an additional change. This is why there will
> not be any second version of that previous patch.
> 
>  fs/exec.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/exec.c b/fs/exec.c
> index 4a2129c0d422..5fa652ca5823 100644
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -639,11 +639,11 @@ int copy_string_kernel(const char *arg, struct linux_binprm *bprm)
>  		page = get_arg_page(bprm, pos, 1);
>  		if (!page)
>  			return -E2BIG;
> -		kaddr = kmap_atomic(page);
> +		kaddr = kmap_local_page(page);
>  		flush_arg_page(bprm, pos & PAGE_MASK, page);

I really question why we can't use memcpy_to_page() here and move the
flush_arg_page() prior to the mapping?

flush_arg_page() only calls flush_cache_page() which does not need the
mapping to work correctly AFAICT.

Ira

>  		memcpy(kaddr + offset_in_page(pos), arg, bytes_to_copy);
>  		flush_dcache_page(page);
> -		kunmap_atomic(kaddr);
> +		kunmap_local(kaddr);
>  		put_arg_page(page);
>  	}
>  
> -- 
> 2.36.1
> 

Re: [PATCH] fs: Call kmap_local_page() in copy_string_kernel()

[Fabio M. De Francesco]

On venerdì 22 luglio 2022 02:14:20 CEST Ira Weiny wrote:
> On Sun, Jul 10, 2022 at 12:01:36PM +0200, Fabio M. De Francesco wrote:
> > The use of kmap_atomic() is being deprecated in favor of 
kmap_local_page().
> > 
> > With kmap_local_page(), the mappings are per thread, CPU local, not
> > globally visible and can take page faults. Furthermore, the mappings 
can be
> > acquired from any context (including interrupts).
> > 
> > Therefore, use kmap_local_page() in copy_string_kernel() instead of
> > kmap_atomic().
> > 
> > Tested with xfstests on a QEMU + KVM 32-bits VM booting a kernel with
> > HIGHMEM64GB enabled.
> > 
> > Suggested-by: Ira Weiny <ira.weiny@intel.com>
> > Signed-off-by: Fabio M. De Francesco <fmdefrancesco@gmail.com>
> > ---
> > 
> > I sent a first patch to fs/exec.c for converting kmap() and 
kmap_atomic()
> > to kmap_local_page():
> > https://lore.kernel.org/lkml/20220630163527.9776-1-fmdefrancesco@gmail.com/
> > 
> > Some days ago, Ira Weiny, while he was reviewing that patch, made me 
notice
> > that I had overlooked a second kmap_atomic() in the same file (thanks):
> > https://lore.kernel.org/lkml/YsiQptk19txHrG4c@iweiny-desk3/
> > 
> > I've been asked to send this as an additional change. This is why there 
will
> > not be any second version of that previous patch.
> > 
> >  fs/exec.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> > 
> > diff --git a/fs/exec.c b/fs/exec.c
> > index 4a2129c0d422..5fa652ca5823 100644
> > --- a/fs/exec.c
> > +++ b/fs/exec.c
> > @@ -639,11 +639,11 @@ int copy_string_kernel(const char *arg, struct 
linux_binprm *bprm)
> >  		page = get_arg_page(bprm, pos, 1);
> >  		if (!page)
> >  			return -E2BIG;
> > -		kaddr = kmap_atomic(page);
> > +		kaddr = kmap_local_page(page);
> >  		flush_arg_page(bprm, pos & PAGE_MASK, page);
> 
> I really question why we can't use memcpy_to_page() here and move the
> flush_arg_page() prior to the mapping?
> 
> flush_arg_page() only calls flush_cache_page() which does not need the
> mapping to work correctly AFAICT.

You're right here. I'm sorry for being so lazy and not checking that 
flush_arg_page() does not need to be called while the task holds the local 
mapping :-(

In v2 I'll move flush_arg_page() one line above memcpy_to_page().

Thanks for your comment,

Fabio

> 
> Ira
> 
> >  		memcpy(kaddr + offset_in_page(pos), arg, 
bytes_to_copy);
> >  		flush_dcache_page(page);
> > -		kunmap_atomic(kaddr);
> > +		kunmap_local(kaddr);
> >  		put_arg_page(page);
> >  	}
> >  
> > -- 
> > 2.36.1
> > 
>