[PATCH] mm/gup.c: Handle error at earliest for incorrect nr_pages value

[PATCH] mm/gup.c: Handle error at earliest for incorrect nr_pages value

[Souptick Joarder]

As per documentation, pin_user_pages_fast() & get_user_pages_fast()
will return 0, if nr_pages <= 0. But this can be figure out only after
going inside the internal_get_user_pages_fast().

This can be handled early. Adding a check for the same.

Signed-off-by: Souptick Joarder <jrdr.linux@gmail.com>
---
 mm/gup.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/mm/gup.c b/mm/gup.c
index 50681f0..a13aaa6 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -2817,6 +2817,8 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
 	 */
 	if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
 		return -EINVAL;
+	if (nr_pages <= 0)
+		return 0;
 
 	/*
 	 * The caller may or may not have explicitly set FOLL_GET; either way is
@@ -2854,6 +2856,8 @@ int pin_user_pages_fast(unsigned long start, int nr_pages,
 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
 	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
 		return -EINVAL;
+	if (nr_pages <= 0)
+		return 0;
 
 	gup_flags |= FOLL_PIN;
 	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
-- 
1.9.1

Re: [PATCH] mm/gup.c: Handle error at earliest for incorrect nr_pages value

[Andrew Morton]

On Fri,  1 May 2020 01:41:58 +0530 Souptick Joarder <jrdr.linux@gmail.com> wrote:

> As per documentation, pin_user_pages_fast() & get_user_pages_fast()
> will return 0, if nr_pages <= 0. But this can be figure out only after
> going inside the internal_get_user_pages_fast().
> 
> This can be handled early. Adding a check for the same.

Calling get_user_pages() for (nrpages <= 0) is a nonsensical thing to
do so it isn't a thing we should care to optimize.  Adding new checks
will actually slow down the use cases which we *do* care about.  And it
adds more code.

Re: [PATCH] mm/gup.c: Handle error at earliest for incorrect nr_pages value

[Ira Weiny]

On Fri, May 01, 2020 at 01:41:58AM +0530, Souptick Joarder wrote:
> As per documentation, pin_user_pages_fast() & get_user_pages_fast()
> will return 0, if nr_pages <= 0. But this can be figure out only after
> going inside the internal_get_user_pages_fast().

Why is nr_pages not unsigned?  I seem to have convinced myself before that
there was a good reason for it but really what is the point of calling either
of these functions with nr_pages not > 0?

> 
> This can be handled early. Adding a check for the same.
> 
> Signed-off-by: Souptick Joarder <jrdr.linux@gmail.com>
> ---
>  mm/gup.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index 50681f0..a13aaa6 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -2817,6 +2817,8 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
>  	 */
>  	if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
>  		return -EINVAL;
> +	if (nr_pages <= 0)
> +		return 0;

I think the documentation may be wrong here...  Is there a caller who expects a
return of 0 for this behavior?

It seems like these should be a warn on and return -EINVAL.  I just don't see
the use case here.

Ira

>  
>  	/*
>  	 * The caller may or may not have explicitly set FOLL_GET; either way is
> @@ -2854,6 +2856,8 @@ int pin_user_pages_fast(unsigned long start, int nr_pages,
>  	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
>  	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
>  		return -EINVAL;
> +	if (nr_pages <= 0)
> +		return 0;
>  
>  	gup_flags |= FOLL_PIN;
>  	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
> -- 
> 1.9.1
> 
> 

Re: [PATCH] mm/gup.c: Handle error at earliest for incorrect nr_pages value

[Souptick Joarder]

On Fri, May 1, 2020 at 2:54 AM Ira Weiny <ira.weiny@intel.com> wrote:
>
> On Fri, May 01, 2020 at 01:41:58AM +0530, Souptick Joarder wrote:
> > As per documentation, pin_user_pages_fast() & get_user_pages_fast()
> > will return 0, if nr_pages <= 0. But this can be figure out only after
> > going inside the internal_get_user_pages_fast().
>
> Why is nr_pages not unsigned?

Not sure of why but it can be unsigned.

>
> >
> > This can be handled early. Adding a check for the same.
> >
> > Signed-off-by: Souptick Joarder <jrdr.linux@gmail.com>
> > ---
> >  mm/gup.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> >
> > diff --git a/mm/gup.c b/mm/gup.c
> > index 50681f0..a13aaa6 100644
> > --- a/mm/gup.c
> > +++ b/mm/gup.c
> > @@ -2817,6 +2817,8 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
> >        */
> >       if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
> >               return -EINVAL;
> > +     if (nr_pages <= 0)
> > +             return 0;
>
> I think the documentation may be wrong here...  Is there a caller who expects a
> return of 0 for this behavior?
>

My understanding is -
{get/pin}_user_pages_fast() -> internal_get_user_pages_fast()

Inside internal_get_user_pages_fast()
...
start = untagged_addr(start) & PAGE_MASK;
addr = start;
len = (unsigned long) nr_pages << PAGE_SHIFT;
end = start + len;

if (end <= start)
return 0;
...
This is the only place where {get/pin}_user_pages_fast() returns 0 which
indirectly checks for nr_pages <= 0. For each call to
{get/pin}_user_pages_fast()
we end up checking *nr_pages <= 0* anyway.

There are some instances where caller of these APIs handles return
value 0 as well. Example -
arch/ia64/kernel/err_inject.c#L145
arch/powerpc/kvm/book3s_64_mmu_hv.c#L582
arch/powerpc/kvm/book3s_64_mmu_hv.c#L1174
drivers/staging/gasket/gasket_page_table.c#L489
drivers/rapidio/devices/rio_mport_cdev.c#L865
drivers/platform/goldfish/goldfish_pipe.c#L277
drivers/gpu/drm/via/via_dmablit.c#L242
(some of the error handling looks incorrect)

If we end up checking *nr_pages <= 0* inside  internal_get_user_pages_fast(),
then why not at first place like this patch does ?

2nd opinion is, inside internal_get_user_pages_fast(), for *nr_pages
<= 0* better
to return -ERRNO rather than 0 ?

Re: [PATCH] mm/gup.c: Handle error at earliest for incorrect nr_pages value

[John Hubbard]

On 2020-04-30 13:11, Souptick Joarder wrote:
> As per documentation, pin_user_pages_fast() & get_user_pages_fast()
> will return 0, if nr_pages <= 0. But this can be figure out only after
> going inside the internal_get_user_pages_fast().
> 
> This can be handled early. Adding a check for the same.


Please, no. For precisely the reasons that Andrew gave: you are
attempting to optimize for a case that doesn't matter other than
for error handling, and which is already handled adequately. And
as he also pointed out, it very slightly UN-optimizes the path that
we *do* care about. So why are you still advocating for this?

If you want to change the gup/pup API so that asking for zero pages
means -EINVAL, then fine, go for it. That's a large thing, and a
tree-wide audit, but if you feel it's worth pursuing then it's at
least consistent.

But this patch here needs to be abandoned.


thanks,
-- 
John Hubbard
NVIDIA

> 
> Signed-off-by: Souptick Joarder <jrdr.linux@gmail.com>
> ---
>   mm/gup.c | 4 ++++
>   1 file changed, 4 insertions(+)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index 50681f0..a13aaa6 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -2817,6 +2817,8 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
>   	 */
>   	if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
>   		return -EINVAL;
> +	if (nr_pages <= 0)
> +		return 0;
>   
>   	/*
>   	 * The caller may or may not have explicitly set FOLL_GET; either way is
> @@ -2854,6 +2856,8 @@ int pin_user_pages_fast(unsigned long start, int nr_pages,
>   	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
>   	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
>   		return -EINVAL;
> +	if (nr_pages <= 0)
> +		return 0;
>   
>   	gup_flags |= FOLL_PIN;
>   	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
> 

Re: [PATCH] mm/gup.c: Handle error at earliest for incorrect nr_pages value

[Souptick Joarder]

On Sat, May 2, 2020 at 3:03 AM John Hubbard <jhubbard@nvidia.com> wrote:
>
> On 2020-04-30 13:11, Souptick Joarder wrote:
> > As per documentation, pin_user_pages_fast() & get_user_pages_fast()
> > will return 0, if nr_pages <= 0. But this can be figure out only after
> > going inside the internal_get_user_pages_fast().
> >
> > This can be handled early. Adding a check for the same.
>
>
> Please, no. For precisely the reasons that Andrew gave: you are
> attempting to optimize for a case that doesn't matter other than
> for error handling, and which is already handled adequately. And
> as he also pointed out, it very slightly UN-optimizes the path that
> we *do* care about. So why are you still advocating for this?
>
> If you want to change the gup/pup API so that asking for zero pages
> means -EINVAL, then fine, go for it. That's a large thing, and a
> tree-wide audit, but if you feel it's worth pursuing then it's at
> least consistent.
>
> But this patch here needs to be abandoned.

Sure, will drop this patch.
>
>
> thanks,
> --
> John Hubbard
> NVIDIA
>
> >
> > Signed-off-by: Souptick Joarder <jrdr.linux@gmail.com>
> > ---
> >   mm/gup.c | 4 ++++
> >   1 file changed, 4 insertions(+)
> >
> > diff --git a/mm/gup.c b/mm/gup.c
> > index 50681f0..a13aaa6 100644
> > --- a/mm/gup.c
> > +++ b/mm/gup.c
> > @@ -2817,6 +2817,8 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
> >        */
> >       if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
> >               return -EINVAL;
> > +     if (nr_pages <= 0)
> > +             return 0;
> >
> >       /*
> >        * The caller may or may not have explicitly set FOLL_GET; either way is
> > @@ -2854,6 +2856,8 @@ int pin_user_pages_fast(unsigned long start, int nr_pages,
> >       /* FOLL_GET and FOLL_PIN are mutually exclusive. */
> >       if (WARN_ON_ONCE(gup_flags & FOLL_GET))
> >               return -EINVAL;
> > +     if (nr_pages <= 0)
> > +             return 0;
> >
> >       gup_flags |= FOLL_PIN;
> >       return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
> >
>